LLM Engines
A reference list of LLM and embedding models by provider: generative models, reasoners, embeddings, and rerankers. Each row lists modalities, parameters, API or repo id, use cases, deployment, pricing, benchmark metrics where public numbers exist (vendor papers, MTEB, or common eval suites), and coarse Scale.
Scale is capacity class: tiny, medium, large; top vendor lines also use large / frontier; teacher–student distillations use tiny / distilled, medium / distilled, or large / distilled matching backbone size (details in prose where it matters). Economical embedding tiers use tiny / medium; strong API defaults use large; flagship lines (e.g. OpenAI text-embedding-3-large, Cohere embed-v4, Voyage voyage-3-large / voyage-code-3, Gemini Embedding 2) use large / frontier. Rerankers and niche tools often stay N/A.
- text-embedding-3-smallOpen AIEmbedding
Released25 Jan 2024
Best forRAG, semantic search, classification, recommendations, clustering, near-duplicate detection
Price~$0.02/1M tokens
DeploymentAPI
ModalitiesText-only
Parameters1,536 dims, 8,191 tokens
Prosstrong MTEB scores, dimension reduction support
ConsAPI-only (no local), text-only
Modeltext-embedding-3-small
Scalemedium
BenchmarkingMTEB English average 62.3% (OpenAI-reported at v3 launch vs 61.0% ada-002). Public HF leaderboard reruns can differ.
- text-embedding-3-largeOpen AIEmbedding
Released25 Jan 2024
Best forRAG, semantic search, classification, recommendations, multilingual search, complex queries
Price~$0.13/1M tokens
DeploymentAPI
ModalitiesText-only
Parameters3,072 dims, 8,191 tokens
ProsHigher quality than 3-small, strong MTEB scores, dimension reduction support
ConsAPI-only (no local), higher cost (~$0.13/1M tokens), text-only
Modeltext-embedding-3-large
Scalelarge / frontier
BenchmarkingMTEB English average 64.6% (OpenAI-reported at v3 launch). Public HF leaderboard reruns can differ.
- CLIP (OpenAI / open source)Open AIEmbedding
Released5 Jan 2021
Best forText–image retrieval, image search by text, zero-shot image classification, cross-modal similarity
PriceFree (open source)
DeploymentLocal (Hugging Face, PyTorch)
ModalitiesText + image
ParametersDepends on variant (ViT-B/32, ViT-L/14, etc.); ~224×224 images; 512–768 dims
ProsOpen source, runs locally, strong text–image alignment, widely used
ConsText + image only (no video/audio), older than newer multimodal models
Modelopenai/clip-vit-base-patch32, laion/CLIP-ViT-L-14, etc.
ScaleN/A
Benchmarking- ImageNet zero-shot (ViT-B/32 — Radford et al. 2021)63.2%
- CIFAR-10 zero-shot (same table)91.3%
- STL-10 zero-shot (same table)97.2%
- Food-101 zero-shot (same table)84.4%
- Oxford-IIIT Pets zero-shot (same table)87.0%
- CIFAR-100 zero-shot (same table)65.1%
- UCF101 zero-shot (same table)64.5%
Zero-shot classification (ViT-B/32): full 27-dataset suite in Radford et al. (2021) Table 11; below are selected columns from that row.
- GPT-6 Astra (OpenAI)Open AIReasoning / thinker
Released3 Sep 2026
Best forHardest long-horizon agent work — terminal / coding agents, computer and browser use, frontier math, cybersecurity research; top tier when cost is secondary. Unique: OpenAI's GPT-6 flagship — saturates FrontierMath Tier 4 (97.6%), ARC-AGI-3 (99.9%) and ExploitBench (100%), and sets a new computer-use frontier (OSWorld 2.0 72.6%).
Price~$10 / 1M input, ~$50 / 1M output
DeploymentAPI (Responses API, SDKs); ChatGPT; Codex
ModalitiesText + image (input)
Parameters~1.05M context; reasoning effort levels; GPT-6.1 Astra was cancelled after alignment regressions, so Astra stays the GPT-6 flagship
ProsStrongest OpenAI model; big jump on Terminal-Bench 4.0; frontier computer use; saturates several math / reasoning benches
ConsMost expensive OpenAI tier (~5× GPT-6.1 Sol); coding roughly on par with Claude Fable 5.1; GPT-6.1 Sol nearly matches it on agentic coding at 1/5 the price
Modelgpt-6-astra
Scalelarge / frontier
BenchmarkingSource: OpenAI-reported (GPT-6 Astra launch) via third-party coverage. Terminal-Bench 4.0 57.9% vs GPT-5.6 Sol 37.3%; FrontierMath Tier 4 97.6%; ARC-AGI-3 99.9%; ExploitBench 100%. Not a third-party rerun.
- GPT-6.1 Sol (OpenAI)Open AIReasoning / thinker
Released29 Sep 2026
Best forDefault production tier for agentic coding, computer use and professional work; Codex / coding-agent backend at scale. Unique: nearly matches GPT-6 Astra on agentic coding at 1/5 of Astra's price; replaces GPT-6 Sol at the same price.
Price~$2 / 1M input, ~$0.10 / 1M cached input, ~$10 / 1M output
DeploymentAPI (Responses API, SDKs); ChatGPT Work; Codex; OpenRouter; Vercel AI Gateway; GitHub Copilot Pro+
ModalitiesText + image (input)
Parameters1,050,000-token context (above 272K billed at 2× input / 1.5× output); 128K max output; reasoning effort low / medium (default) / high / xhigh / max
ProsNear-Astra agentic coding at $2 / $10; very cheap cached input; broad availability
ConsBelow Astra on hard troubleshooting / exploit tasks; long-context surcharge above 272K; no none / minimal reasoning
Modelgpt-6.1-sol
Scalelarge / frontier
BenchmarkingSource: OpenAI-reported / third-party (DataCamp, Vals.ai). DeepSWE v1.1 ≈ GPT-6 Astra and +6.4 pt vs GPT-6 Sol; ExploitBench internal port 21.5% (Astra 31.5%); Vals Index 61.15% (#8 of 43).
- GPT-6 Luna (OpenAI)Open AIGenerative
Released22 Sep 2026
Best forHigh-volume, latency-sensitive work — classification, summaries, routing, light coding and agent steps; cheapest GPT-6 tier. Unique: brings GPT-6 Astra advances to the cheapest tier at $0.10 / $0.50.
Price~$0.10 / 1M input, ~$0.50 / 1M output
DeploymentAPI (Responses API, SDKs); ChatGPT (incl. Free / Go in desktop app); Codex
ModalitiesText + image (input)
ParametersGPT-6 family fast tier; ~1.05M context; 128K max output
ProsExtremely cheap; fast; completes tasks GPT-5.6 Luna failed per developer reports
ConsLowest ceiling in GPT-6 family; few published benchmarks; specs not fully documented at launch
Modelgpt-6-luna
Scalemedium
Benchmarking- BenchmarkingN/A
No standard-suite numbers published at launch; OpenAI positions it as faster and cheaper than GPT-5.6 Luna.
- GPT-5.6 Sol (OpenAI)Open AIReasoning / thinker
Released9 Jul 2026
Best forPrevious-generation flagship for long-horizon agent work, cybersecurity, computer use and browsing. Unique: Sol Ultra mode runs 4 parallel agents (~3× cost) — Terminal-Bench 2.1 91.9% vs 88.8% single-agent.
PriceList ~$5 / 1M input, ~$30 / 1M output; promo ~$4 / ~$20 (through at least 21 Nov 2026)
DeploymentAPI (Responses API, SDKs); ChatGPT; Codex
ModalitiesText + image (input)
ParametersReasoning effort none → max; Sol Ultra = 4 parallel agents; bare gpt-5.6 alias routes to Sol
ProsStrong agentic benches; strong long-context recall; big cybersecurity jump vs GPT-5.5
ConsSuperseded by GPT-6 Astra / GPT-6.1 Sol (cheaper and better); expensive list price; behind Claude Fable 5 on SWE-Bench Pro
Modelgpt-5.6-sol (alias: gpt-5.6)
Scalelarge / frontier
BenchmarkingSource: OpenAI-reported (GPT-5.6 launch) via third-party analysis. Also: SWE-Bench Pro 64.6%; MRCR 91.5%; Agents' Last Exam 53.6; AA Intelligence Index v4.1: 59.
- GPT-5.6 Terra (OpenAI)Open AIReasoning / thinker
Released9 Jul 2026
Best forBalanced everyday default in the GPT-5.6 family — coding, agents, analysis at half Sol's price. Unique: within ~1–3 pt of Sol on most evals while costing half.
PriceList ~$2.50 / 1M input, ~$15 / 1M output; promo ~$2 / ~$12
DeploymentAPI (Responses API, SDKs); ChatGPT; Codex
ModalitiesText + image (input)
ParametersGPT-5.6 mid tier; reasoning effort levels as Sol
ProsNear-Sol quality at half price; good routing default
ConsSuperseded by GPT-6 family; no Ultra mode
Modelgpt-5.6-terra
Scalelarge
BenchmarkingSource: OpenAI-reported (GPT-5.6 launch) via third-party analysis. Also: Terminal-Bench 2.1 ~84–85%; Agents' Last Exam ~50.4.
- GPT-5.6 Luna (OpenAI)Open AIGenerative
Released9 Jul 2026
Best forHigh-volume classification, summaries and lightweight agent steps in the GPT-5.6 family. Unique: ~1/5 of Sol's price while staying within ~3 pt on Agents' Last Exam.
PriceList ~$1 / 1M input, ~$6 / 1M output; promo ~$0.20 / ~$1.20
DeploymentAPI (Responses API, SDKs); ChatGPT; Codex
ModalitiesText + image (input)
ParametersGPT-5.6 fast tier
ProsCheap; DeepSWE strong for the price
ConsWeak long-context recall — avoid for long documents; superseded by GPT-6 Luna
Modelgpt-5.6-luna
Scalemedium
BenchmarkingSource: OpenAI-reported (GPT-5.6 launch) via third-party analysis. Also: Terminal-Bench 2.1 ~80–83%; Agents' Last Exam ~50.3.
- BGE-large-en-v1.5 (BAAI), BAAI General Embedding, BAAI = Beijing Academy of Artificial IntelligenceSentence Transformers (Hugging Face)Embedding
ReleasedSep 2023
Best forRAG, semantic search, retrieval, classification, open-source deployments
PriceFree (open source)
DeploymentLocal
ModalitiesText-only
Parameters1,024 dims, 512 tokens, ~335M params
ProsStrong MTEB scores, open source, runs locally, good quality/size trade-off
ConsText-only, shorter context than some models
ModelBAAI/bge-large-en-v1.5
Scalelarge
BenchmarkingMTEB (English, v1): ~64.2%
- GTE-large-en-v1.5 (Alibaba) - General Text EmbeddingsSentence Transformers (Hugging Face)Embedding
ReleasedApr 2024
Best forRAG, search, classification, English long-context
PriceFree (open source)
DeploymentLocal
ModalitiesText-only
Parameters1,024 dims, 8,192 tokens, ~335M params
ProsHigh MTEB score, long context
ConsEnglish-only; other GTE checkpoints for multilingual
ModelAlibaba-NLP/gte-large-en-v1.5
Scalelarge
BenchmarkingMTEB: 65.39% (Average 56 tasks; mteb 1.2.0 eval setting on model card).
- E5-mistral-7b-instructSentence Transformers (Hugging Face)Embedding
ReleasedJan 2024 (paper 31 Dec 2023)
Best forHigh-quality retrieval, RAG, complex queries
PriceFree (open source)
DeploymentLocal
ModalitiesText-only
Parameters4,096 dims, 32K tokens, 7B params
ProsTop MTEB performance, long context
ConsHeavy; needs more GPU memory
Modelintfloat/e5-mistral-7b-instruct
Scalemedium
BenchmarkingMTEB English (56 datasets) mean 66.6%; Wang et al. (2024) Table 2.
- BGE-reranker-largeSentence Transformers (Hugging Face)Reranker
ReleasedSep 2023
Best forRAG reranking, passage–query relevance
PriceFree (open source)
DeploymentLocal
ModalitiesText-only
ParametersCross-encoder, query + passage → score
ProsStrong MTEB, works with Sentence Transformers
ConsText-only
ModelBAAI/bge-reranker-large
ScaleN/A
Benchmarking- MTEB T2 reranking (MAP)67.6
- MTEB T2 ZH→EN (MAP)64.03
- MTEB T2 EN→ZH (MAP)61.44
- MTEB MMarco reranking (MAP)37.16
- MTEB CMedQAv1 (MAP)82.15
- MTEB CMedQAv2 (MAP)84.18
- C-MTEB rerank mean (MAP)66.09
C-MTEB reranking tasks, MAP (×100), from BAAI Hugging Face model card table.
- BGE-reranker-v2-m3Sentence Transformers (Hugging Face)Reranker
ReleasedMar 2024
Best forMultilingual reranking
PriceFree
DeploymentLocal
ModalitiesText-only (multilingual)
ParametersMultilingual cross-encoder
Pros100+ languages
ConsHeavier than base
ModelBAAI/bge-reranker-v2-m3
ScaleN/A
Benchmarking- BEIR Q&A mean (NDCG@10)0.6734
- BEIR NQ (NDCG@10)0.7028
- BEIR HotpotQA (NDCG@10)0.8635
- BEIR FiQA (NDCG@10, NV-Mistral7B-v2 pool)0.4539
- BEIR FiQA NDCG@10 (bge-large-en-v1.5 pool)44.83
- BEIR FiQA NDCG@100 (bge-large-en-v1.5 pool)51.53
- BEIR FiQA MRR@10 (bge-large-en-v1.5 pool)53.38
Table 1 (arXiv:2409.07691): NDCG@10 as in paper (0–1). Row + bge-reranker-v2-m3 under Embedding: NV-EmbedQA-Mistral7B-v2 (top-100 rerank). FiQA bge-large pool: README figure.
- Gemini Embedding 2GoogleEmbedding
Released10 Mar 2026 (preview); GA 22 Apr 2026
Best forMultimodal RAG, cross-modal search, semantic search over text, images, video, audio, documents
Pricetext ~$0.20/1M tokens, images ~$0.45/1M tokens, audio ~$6.50/1M tokens, video ~$12.00/1M tokens
DeploymentAPI
ModalitiesText, image, video, audio, documents
Parameters3,072 dims, 8K tokens (text), images (up to 6), video (~120s), audio (~80s)
ProsNative multimodal, single embedding space, Matryoshka-style compression, 100+ languages
ConsAPI-only (no local), higher cost than text-only models
Modelgemini-embedding-2-preview
Scalelarge / frontier
Benchmarking- MMTEB task mean68.32
- MTEB (English, v2) task mean73.30
- MTEB (Code)74.66
- XOR-Retrieve Recall@590.42
- XTREME-UP MRR@1064.33
- Borda rank MMTEB (multilingual)#1
Source: Gemini Embedding technical report arXiv:2503.07891 / Google AI (vendor-reported).
- Gemini Embedding 2 (Vertex AI)GoogleEmbedding
Released10 Mar 2026 (preview); GA 22 Apr 2026
Best forMultimodal RAG, cross-modal search, semantic search over text, images, video, audio, documents
Pricetext: ~$0.15 / 1M tokens (cheaper than Gemini API), images/audio/video: similar to Gemini API
DeploymentAPI (Vertex AI)
ModalitiesText, image, video, audio, documents
Parameters3,072 dims, 8K tokens (text), images (up to 6), video (~120s), audio (~80s) · Vertex text embedding ~$0.15/1M vs Gemini API ~$0.20/1M is typical; confirm on current pricing pages.
ProsNative multimodal, single embedding space, Matryoshka-style compression, 100+ languages, enterprise support, VPC integration
ConsAPI-only (no local), higher cost than text-only models
ModelVertex AI (text-multilingual-embedding-002 or gemini-embedding-2)
Scalelarge / frontier
Benchmarking- MMTEB task mean68.32
- MTEB (English, v2) task mean73.30
- MTEB (Code)74.66
- XOR-Retrieve Recall@590.42
- XTREME-UP MRR@1064.33
- Borda rank MMTEB (multilingual)#1
Same Gemini Embedding family weights as Gemini API; benchmarks from Gemini Embedding technical report (vendor-reported).
- Gemini 2.5 Pro (Thinking mode)GoogleReasoning / thinker
Released25 Mar 2025 (experimental); GA 17 Jun 2025
Best forMath, coding, multi-step reasoning, agents
Price~$1.25/1M in, $10/1M out (prompt ≤200k tokens); ~$2.50/1M in, $15/1M out (>200k)
DeploymentGemini API, Vertex AI
ModalitiesText, image, audio, video
ParametersExtended reasoning, 1M context
ProsStrong reasoning, multimodal
ConsSlower, higher cost
Modelgemini-2.5-pro (with thinking enabled)
Scalelarge / frontier
Benchmarking- GPQA Diamond (pass@1)86.4%
- SWE-Bench Verified (single attempt)59.6%
- MMMU (pass@1)82.0%
- VideoMMMU83.6%
- Global MMLU89.2%
- AIME 2025 (pass@1)88.0%
Source: Gemini 2.5 Pro Model Card / Gemini 2.5 tech report (Google DeepMind; vendor-reported pass@1 unless noted).
- Gemini 3.1 Pro PreviewGoogleReasoning / thinker
Released19 Feb 2026
Best forComplex reasoning, coding, agents, multimodal analysis
PricePaid (split by prompt length): input $2.00/1M (≤200k tokens) or $4.00/1M (>200k); output $12.00/1M or $18.00/1M
DeploymentGemini API, Vertex AI
ModalitiesText, image, audio, video (per model doc)
ParametersSee model doc (preview; limits updated there)
ProsNewest Pro tier in Gemini 3.1 line, strong agentic / coding positioning
ConsPreview (behavior/rates may change); higher cost than Flash
Modelgemini-3.1-pro-preview
Scalelarge / frontier
Benchmarking- GPQA Diamond94.3%
- SWE-Bench Verified (single attempt)80.6%
- SWE-Bench Pro (single attempt)54.2%
- Terminal-Bench 2.068.5%
- ARC-AGI-2 (verified)77.1%
- Humanity's Last Exam (no tools)44.4%
Source: Gemini 3.1 Pro model card (deepmind.google; Feb 2026; Gemini 3.1 Pro Thinking High vs external baselines table).
- Gemini 2.5 Flash (Thinking)GoogleReasoning / thinker
Released17 Apr 2025 (preview); GA 17 Jun 2025
Best forLightweight reasoning, cost-sensitive reasoning
Price~$0.30/1M input, $2.50/1M output (thinking counted in output; audio in $1/1M)
DeploymentGemini API, Vertex AI
ModalitiesText, image, audio, video
ParametersSmaller reasoning model
ProsCheaper than Pro thinking
ConsLess capable than Pro
Modelgemini-2.5-flash-thinking
Scalemedium
Benchmarking- SWE-Bench Verified (single attempt)60.4%
- GPQA Diamond82.8%
- AIME 2025 (no tools)72.0%
- MMMU-Pro (no tools)66.7%
- Terminal-Bench 2.016.9%
- Humanity's Last Exam (no tools)11.0%
Source: Gemini 3 Flash page (deepmind.google) Performance table — Gemini 2.5 Flash Thinking column.
- Gemini 3 Flash Preview (Thinking)GoogleReasoning / thinker
Released17 Dec 2025
Best forFast Gemini 3 Flash tier, reasoning, agents, search/grounding-heavy work
PricePaid (standard): ~$0.50/1M input (text/image/video), ~$1.00/1M (audio); ~$3.00/1M output (thinking in output)
DeploymentGemini API, Vertex AI
ModalitiesText, image, audio, video, PDF
ParametersPreview; 1M in / 65k out; thinking via API config
ProsNewer Flash line than 2.5; strong speed + capability mix
ConsPreview; higher $/1M than 2.5 Flash; stricter limits than stable models
Modelgemini-3-flash-preview
Scalemedium
Benchmarking- SWE-Bench Verified (single attempt)78.0%
- GPQA Diamond90.4%
- MMMU-Pro (no tools)81.2%
- Terminal-Bench 2.0 (Terminus-2)47.6%
- LiveCodeBench Pro (Elo)2316
- Humanity's Last Exam (no tools)33.7%
Source: Gemini 3 Flash model page (deepmind.google/models/gemini/flash; Gemini 3 Flash Thinking column).
- Gemma 4 E2B (Google / DeepMind)GoogleGenerative
Released31 Mar 2026
Best forOn-device and edge chat, agents, coding; multimodal (text, image, video, audio) on small sizes
PriceOpen weights (infra cost); see AI Studio / Vertex if hosted
DeploymentLocal, Hugging Face, Kaggle, Ollama, Vertex AI
ModalitiesText, image, video, audio
ParametersE2B line, PLE architecture, 128K context (see model card for parameterization)
ProsTiny footprint for Gemma 4 line, native multimodal on E2B/E4B, function calling, thinking modes
ConsLower ceiling than 31B / MoE variants
Modelgoogle/gemma-4-E2B-it
Scaletiny
Benchmarking- MMLU Pro60.0%
- LiveCodeBench v644.0%
- GPQA Diamond43.4%
- Tau2 (avg of 3 domains)24.5%
- MMMU Pro44.2%
- MMMLU67.4%
Source: Gemma 4 Model Card README table (ai.google.dev; instruct models).
- Gemma 4 E4B (Google / DeepMind)GoogleGenerative
Released31 Mar 2026
Best forEdge and browser-class workloads, agents, coding; multimodal (text, image, video, audio)
PriceOpen weights (infra cost); see AI Studio / Vertex if hosted
DeploymentLocal, Hugging Face, Kaggle, Ollama, Vertex AI
ModalitiesText, image, video, audio
ParametersE4B line, PLE architecture, 128K context (see model card for parameterization)
ProsStrong quality for size class, multimodal on small Gemma 4, Apache-2.0-style Gemma terms
ConsHigher static VRAM than naive 4B due to PLE tables
Modelgoogle/gemma-4-E4B-it
Scaletiny
Benchmarking- MMLU Pro69.4%
- LiveCodeBench v652.0%
- GPQA Diamond58.6%
- Tau2 (avg of 3 domains)42.2%
- MMMU Pro52.6%
- MMMLU76.6%
Source: Gemma 4 Model Card README table (ai.google.dev; instruct models).
- Gemma 4 26B A4B MoE (Google / DeepMind)GoogleReasoning / thinker
Released31 Mar 2026
Best forHigh-throughput reasoning, coding, agents; ~4B active params per token
PriceOpen weights (infra cost); full expert set loaded at inference
DeploymentLocal, Hugging Face, Kaggle, Vertex AI
ModalitiesText, image, video, audio
Parameters26B MoE (A4B active per token), 256K context, function calling, thinking modes
ProsEfficient per-token cost vs dense 26B-class, strong reasoning positioning
ConsMemory footprint closer to full 26B than to 4B active
Modelgoogle/gemma-4-26B-A4B-it
Scalelarge
Benchmarking- MMLU Pro82.6%
- LiveCodeBench v677.1%
- GPQA Diamond82.3%
- Tau2 (avg of 3 domains)68.2%
- Codeforces Elo1718
- MRCR v2 (8 needle, 128k avg)44.1%
Source: Gemma 4 Model Card README table (ai.google.dev; instruct models).
- Gemma 4 31B (Google / DeepMind)GoogleReasoning / thinker
Released31 Mar 2026
Best forDeep reasoning, coding, enterprise and server-grade open-weight deployment
PriceOpen weights (infra cost)
DeploymentLocal, Hugging Face, Kaggle, Vertex AI
ModalitiesText, image, video, audio
Parameters31B dense, 256K context, multimodal (text, image, video, audio), function calling
ProsTop open-weight Gemma 4 dense tier, long context, agentic tooling
ConsHeavy GPU/TPU requirements at full precision
Modelgoogle/gemma-4-31B-it
Scalelarge / frontier
Benchmarking- MMLU Pro85.2%
- LiveCodeBench v680.0%
- GPQA Diamond84.3%
- AIME 2026 (no tools)89.2%
- Codeforces Elo2150
- MRCR v2 (8 needle, 128k avg)66.4%
Source: Gemma 4 Model Card README table (ai.google.dev; instruct models).
- Vertex AI Ranking APIGoogleReranker
Released9 Apr 2025 (004 models; 005 preview 1 Sep 2026)
Best forRAG reranking, semantic ranking
Price$1.00 USD per 1,000 ranking calls
DeploymentVertex AI
ModalitiesText-only
ParametersSemantic reranker, <100ms latency
ProsLow latency, strong performance
ConsVertex AI only, API-only
Modelsemantic-ranker-default-004, semantic-ranker-fast-004
ScaleN/A
Benchmarking- TREC DL mean (NDCG@10, 512-003, Pinecone)64.89
- Latency (Vertex RAG docs)<100ms typical
- Default-004 vs API baselines (vendor blog)≥2× faster
- Fast-004 vs default-004 (vendor blog)~3× lower latency
- Max tokens / ranking request (vendor blog)200k
- Max tokens / record (004, docs)1024
- Max records / request (docs)200
NDCG@10 as ×100 where applicable. TREC DL merged eval: Pinecone Dec 2024 (https://www.pinecone.io/blog/pinecone-rerank-v0-announcement/) — google-semantic-ranker-512-003 via API, top-200 rerank, table under TREC section — not default-004, not BEIR. Default-004 BEIR: Google Cloud blog May 30, 2025 (Fig. 2 vs APIs) + GCP ranking_api_beir_evaluation.ipynb. Limits/latency: blog + Vertex Ranking docs.
- Gemini 4 Argon (Google)GoogleReasoning / thinker
Released30 Sep 2026 (limited, Fairwind cyber partners); GA 13 Oct 2026
Best forHardest professional and engineering work — coding, debugging and codebase migrations, defensive cybersecurity (find, validate and patch vulnerabilities), finance / legal / tax analysis, deep research, long-horizon reasoning over text, images, video and files. Unique: Google's first Gemini 4 model and #1 on the Vals Index (68.90%, 1 of 48) ahead of GPT-6 Astra, Claude Fable and Opus; trained specifically for defensive cyber work; 1M unified input-output token limit.
PriceIntro (to 30 Nov 2026) ~$2 / 1M input, ~$10 / 1M output, cached input 95% off; then ~$4 / ~$20
DeploymentGemini API, Google AI Studio, Android Studio, Vertex / Gemini for Business, Google Antigravity; Gemini app and AI Mode in Search (AI Ultra / Pro); Fairwind Program for cyber partners
ModalitiesText + image + video + file in, text out
ParametersParameter count / architecture not published; 1M-token unified context (input + output); max output 262K; multimodal input (text, image, video, file), text output; thinking levels not documented at launch; safety stack with real-time agent-misalignment monitoring (Gray Swan prompt-injection attack success 0.7%)
ProsTop Vals Index (68.90%, #1/48); #1 on Finance Agent v2 (65.40%) and IOI (100%); strong coding (Vibe Code Bench 91.91%, #2/112) and legal (LegalBench 88.30%); very long 1M context with 262K output; native multimodal input; intro price $2 / $10 with cached input 95% off; available in API, AI Studio, Antigravity, Gemini app
ConsClosed weights; price doubles to $4 / $20 after 30 Nov 2026; very slow on long evals (Vals Index run ~46 min); weak on computer-use (CUA-bench 4.83%, 8/9); Google's own launch post gives no head-to-head numbers vs Astra / Fable; consumer access only for Google AI Ultra / Pro subscribers; rest of Gemini 4 family not out yet
Modelgemini-4-argon
Scalelarge / frontier
BenchmarkingVals Index 68.90% (#1/48); Finance Agent v2 65.40% (#1/78); IOI 100% (#1/44); Vibe Code Bench 91.91% (#2/112); LegalBench 88.30% (#3/150); CyberBench v1.1 77.86% (#3/47); ProofBench v1.1 99.0%; CUA-bench 4.83% (8/9); Google-reported: ERQA (robotics) 87%
- embed-english-v3.0CohereEmbedding
Released2 Nov 2023
Best forEnglish RAG, search, classification
Price~$0.10/1M tokens
DeploymentAPI
ModalitiesText + image
Parameters1,024 dims, English; API supports text and images
ProsStrong search, RAG
ConsEnglish only
Modelembed-english-v3.0
Scalelarge
Benchmarking- MTEB NQ (NDCG@10)61.556
- MTEB HotpotQA (NDCG@10)70.723
- MTEB FiQA2018 (NDCG@10)42.189
- MTEB QuoraRetrieval (NDCG@10)88.722
- MTEB MSMARCO (NDCG@10)42.931
- MTEB FEVER (NDCG@10)88.966
- MTEB ArguAna (NDCG@10)61.521
NDCG@10 as ×100 (percentage points). Values match the public Hugging Face model card model-index for this checkpoint. MTEB leaderboard: https://huggingface.co/spaces/mteb/leaderboard — Published MTEB result files (dataset tree): https://huggingface.co/datasets/mteb/results/tree/main/results/Cohere__Cohere-embed-english-v3.0 — Model card: https://huggingface.co/CohereLabs/Cohere-embed-english-v3.0
- embed-v4CohereEmbedding
Released15 Apr 2025
Best forMultimodal RAG, text + image + PDF, mixed content
Price~$0.10–0.12/1M tokens
DeploymentAPI
ModalitiesText + image
ParametersMultimodal (text, image), unified embedding space
ProsMultimodal, high-res images, PDFs
ConsAPI-only
Modelembed-v4
Scalelarge / frontier
Benchmarking- MTEB avg (reported rollup)69.8%
- Retrieval subset (reported)58.2%
- Classification subset (reported)78.4%
- MultimodalText + image in shared space
69.8 / 58.2 / 78.4 appear in the comparison table on https://getathenic.com/blog/cohere-embed-v4-multilingual-embeddings (third-party; links Cohere’s announcement). Cohere’s own post does not surface those three numbers as searchable text: https://cohere.com/blog/embed-4 — Per-task MTEB JSONs: https://huggingface.co/datasets/mteb/results/tree/main/results/Cohere__Cohere-embed-v4.0/1
- embed-multilingual-v3.0CohereEmbedding
Released2 Nov 2023
Best forMultilingual RAG, search
Price~$0.10/1M tokens
DeploymentAPI
ModalitiesText + image
Parameters1,024 dims, 100+ languages; API supports text and images
ProsMultilingual
ConsAPI-only
Modelembed-multilingual-v3.0
Scalelarge
Benchmarking- MTEB (multilingual cluster, typical)~61–64
- Languages (API)100+
- vs Multilingual E5 (Voyage suite cited)+3.89%
- vs OpenAI v3 large (Voyage suite cited)+4.55%
Source: Voyage multilingual comparison (Voyage v3 blog cites +3.89% vs multilingual-e5-large on multilingual suite; CMv3 third-party MTEB clusters ~60–64).
- Command RCohereGenerative
Released11 Mar 2024 (08-2024 refresh: 30 Aug 2024)
Best forRAG, chat, cost-sensitive use
Price~$0.15/1M input, $0.60/1M output
DeploymentAPI
ModalitiesText-only
Parameters128K context
ProsCheaper than R+, solid RAG
ConsLess capable than R+
Modelcommand-r
Scalelarge
Benchmarkingcommand-r-08-2024 generation; Hugging Face weights README is login-gated. https://docs.cohere.com/docs/command-r — https://docs.cohere.com/changelog/command-gets-refreshed — https://cohere.com/blog/command-series-0824 — https://huggingface.co/datasets/open-llm-leaderboard/contents — https://huggingface.co/docs/leaderboards/open_llm_leaderboard/archive — https://huggingface.co/datasets/open-llm-leaderboard/results
- Command R+CohereGenerative
Released4 Apr 2024
Best forRAG, tool use, agents, multilingual
Price~$2.50/1M input, $10/1M output
DeploymentAPI
ModalitiesText-only
Parameters128K context, 23 languages
ProsStrong RAG, tool use, multilingual
ConsHigher cost than Command R
Modelcommand-r-plus
Scalelarge
Benchmarking- MMLU (0-shot, reported)75.7%
- GSM8K (8-shot, reported)70.7%
- HumanEval (pass@1, reported)~79%
- Open LLM leaderboard (avg, reported)74.6%
command-r-plus-08-2024 generation; open-weights HF card is often login-gated. https://docs.cohere.com/docs/command-r-plus — https://docs.cohere.com/changelog/command-gets-refreshed — https://cohere.com/blog/command-series-0824 — https://huggingface.co/datasets/open-llm-leaderboard/contents — https://huggingface.co/docs/leaderboards/open_llm_leaderboard/archive — https://huggingface.co/datasets/open-llm-leaderboard/results
- Command R7BCohereGenerative
Released13 Dec 2024
Best forHigh volume, simple tasks
Price~$0.04/1M input, $0.15/1M output
DeploymentAPI
ModalitiesText-only
Parameters7B params
ProsFast, cheap
ConsLess capable
Modelcommand-r7b
Scaletiny
Benchmarking- MMLU (typical small-command tier)~55–60%
- HumanEval (typical)~65–75%
- RoleHigh-throughput routing / simple RAG
command-r7b-12-2024. https://docs.cohere.com/docs/command-r7b — https://cohere.com/blog/command-r7b — https://huggingface.co/datasets/open-llm-leaderboard/contents — https://huggingface.co/docs/leaderboards/open_llm_leaderboard/archive — open weights: https://huggingface.co/CohereForAI/c4ai-command-r7b-12-2024
- Command A ReasoningCohereReasoning / thinker
Released21 Aug 2025
Best forAgents, tool use, complex reasoning
PriceHigher than Command R+
DeploymentAPI
ModalitiesText-only
Parameters111B params, 256K context, 23 languages
ProsStrong reasoning, tool use, agentic
ConsAPI-only, higher cost
Modelcommand-a-reasoning-08-2025
Scalelarge / frontier
BenchmarkingScores from Cohere Command A technical report (arxiv:2504.00698, Tables 1, 3, 15): flagship Command A. Command A Reasoning is the reasoning-tuned 111B variant; treat as same family, not byte-for-byte identical.
- rerank-v4.0-proCohereReranker
Released11 Dec 2025
Best forRAG reranking, high relevance
PricePer search
DeploymentAPI
ModalitiesText-only
Parameters32K context
ProsHigh quality
ConsSlower than fast
Modelrerank-v4.0-pro
ScaleN/A
Benchmarking- T2-RAGBench R@147.2
- T2-RAGBench R@375.8
- T2-RAGBench R@581.6
- T2-RAGBench R@1086.1
- T2-RAGBench MRR@360.5
- T2-RAGBench nDCG@1068.3
- T2-RAGBench MAP62.5
Metrics as ×100 (proportion×100). arXiv:2604.01733 Table I, T2-RAGBench (23,088 queries): hybrid BM25+dense RRF, 50 candidates to Cohere rerank-v4.0-pro, top-10. Pipeline end-to-end on finance text+table retrieval, not reranker-only BEIR.
- rerank-v3.5CohereReranker
Released2 Dec 2024
Best forMultilingual reranking
Price~$2/1,000 searches
DeploymentAPI
ModalitiesText-only (multilingual)
Parameters4,096 tokens
ProsMultilingual
ConsShorter context than v4
Modelrerank-v3.5
ScaleN/A
Benchmarking- Reasoning suite P@1 (2 candidates)81.59
- Multilingual suite mean nDCG@1062.18
- Same suite Hybrid nDCG@1052.10
- Same suite Dense nDCG@1053.83
- Same suite Cohere Rerank 3 nDCG@1052.27
- Reasoning suite Hybrid P@148.80
- Reasoning suite Dense P@150.64
Vendor suite numbers (Cohere, republished Microsoft Foundry catalog for Cohere-rerank-v3.5, Jan 2026). Reasoning: P@1 on adversarial two-candidate queries. Multilingual: mean nDCG@10 on Cohere multilingual suite (18 languages). Baselines are same-suite; not open BEIR/MIRACL tables.
- voyage-3-largeVoyage AIEmbedding
Released7 Jan 2025
Best forHigh-accuracy RAG, semantic search, multilingual retrieval, enterprise retrieval quality
PricePremium tier (check latest Voyage pricing page)
DeploymentAPI
ModalitiesText-only
ParametersText embeddings, long-context input, Matryoshka-style dimension shortening support
ProsVery strong retrieval quality, good multilingual performance, flexible embedding size
ConsAPI-only (no local), higher cost than lite variants
Modelvoyage-3-large
Scalelarge / frontier
Benchmarking- NDCG@10 (est. from voyage-3 76.72 + 4.14% rel.)~79.9
- vs OpenAI v3 large (avg domains, vendor)+9.74%
- vs voyage-3 (avg domains, vendor)+4.14%
- Context length32K
Source: voyage-3-large blog (+4.14% avg vs voyage-3 on eight-domain suite; NDCG@10 baseline from voyage-3 post).
- voyage-3Voyage AIEmbedding
Released18 Sep 2024
Best forGeneral-purpose RAG, semantic search, recommendations, classification
PriceMid tier (check latest Voyage pricing page)
DeploymentAPI
ModalitiesText-only
ParametersText embeddings, long-context input, balanced quality/latency
ProsStrong quality with better cost than large
ConsAPI-only, not as accurate as large on hard retrieval sets
Modelvoyage-3
Scalemedium
BenchmarkingNDCG@10: 76.72%
- voyage-3-liteVoyage AIEmbedding
Released18 Sep 2024
Best forHigh-volume embedding pipelines, cost-sensitive search/RAG, near-duplicate detection
PriceLow-cost tier (check latest Voyage pricing page)
DeploymentAPI
ModalitiesText-only
ParametersLightweight text embeddings, lower latency
ProsCheap, fast, scalable
ConsLower retrieval quality than voyage-3 / voyage-3-large
Modelvoyage-3-lite
Scaletiny
Benchmarking- NDCG@10 (avg suite)72.98
- vs OpenAI v3 large (avg domains, vendor)+3.82%
- vs OpenAI v3 small same price (vendor)+7.58%
- Context length32K
Source: voyage-3 blog (NDCG@10 table; Sept 2024).
- voyage-code-3Voyage AIEmbedding
Released4 Dec 2024
Best forCode search, code RAG, repository retrieval, code similarity
PriceSpecialized tier (check latest Voyage pricing page)
DeploymentAPI
ModalitiesText + code
ParametersCode-focused embeddings for natural language + code retrieval
ProsBetter code retrieval than general text embeddings
ConsAPI-only, less ideal for purely non-code corpora
Modelvoyage-code-3
Scalelarge / frontier
Benchmarking- NDCG@10 @1024 dims (avg code suite)92.28%
- vs OpenAI v3 large avg lift (vendor)+13.80%
- vs CodeSage-large avg lift (vendor)+16.81%
- Context length32K
Source: voyage-code-3 blog (Dec 2024; 32 code retrieval datasets; NDCG@10 at 1024-dim vs baselines).
- rerank-2Voyage AIReranker
Released30 Sep 2024
Best forRe-ranking top-k retrieved documents for higher precision in RAG
PricePer-request/token based (check latest Voyage pricing page)
DeploymentAPI
ModalitiesText-only
ParametersCross-encoder reranker (query-document relevance scoring)
ProsNoticeable precision boost after initial vector retrieval
ConsExtra latency/cost step after retrieval
Modelrerank-2
ScaleN/A
Benchmarking- ICLERB nDCG@1063.86
- ICLERB nDCG@5064.32
- Rel. mean NDCG@10 lift vs text-embedding-3-large (%)13.89
- Mean NDCG@10 − rerank-1 (pp, 3 first-stages)2.84
- Mean NDCG@10 − Cohere rerank-english-v3.0 (pp)6.33
- Mean NDCG@10 − bge-reranker-v2-m3 (pp)14.75
- Mean NDCG@10 − Cohere rerank-multilingual-v3.0 (pp, 51 ds.)8.83
ICLERB nDCG as ×100: arXiv:2411.18947 Table 2 (in-context-learning retrieval; rankings differ from generic RAG). Vendor NDCG@10 suite: blog.voyageai.com/2024/09/30/rerank-2, 93 datasets, top-100→rerank→10, three first-stages (BM25+GTE-large-en-v1.5, text-embedding-3-large, voyage-multilingual-2).
- rerank-2-liteVoyage AIReranker
Released30 Sep 2024
Best forCost-sensitive re-ranking at larger scale
PriceLower than rerank-2 (check latest Voyage pricing page)
DeploymentAPI
ModalitiesText-only
ParametersLightweight reranker optimized for speed/cost
ProsFaster and cheaper than rerank-2
ConsSlightly lower precision than full rerank-2
Modelrerank-2-lite
ScaleN/A
Benchmarking- Rel. mean NDCG@10 lift vs text-embedding-3-large (%)11.86
- Mean rel. lift below rerank-2 (pp, same 93-ds. TL;DR)2.03
- Mean NDCG@10 − Cohere rerank-english-v3.0 (pp)4.49
- Mean NDCG@10 − bge-reranker-v2-m3 (pp)12.91
- Mean NDCG@10 − Cohere rerank-multilingual-v3.0 (pp, 51 ds.)6.24
- Mean NDCG@10 − bge-reranker-v2-m3 (pp, multilingual)2.26
- Rel. lift above Cohere v3 stack lift (pp, vendor intro)5.12
Vendor NDCG@10 suite: blog.voyageai.com/2024/09/30/rerank-2 (same protocol as rerank-2: 93 English-domain + multilingual slices, top-100→rerank→10, three first-stages). No third-party ICLERB row for lite.
- Qwen2.5-72B-Instruct (Alibaba / Qwen)Alibaba QwenGenerative
Released19 Sep 2024
Best forHigh-quality chat, analysis, coding, complex instruction following
PriceOpen weights (inference cost depends on your infra) or provider API pricing
DeploymentLocal or API (provider-dependent)
ModalitiesText-only
Parameters72B open-weight instruct model, long-context variants available
ProsStrong general quality, good multilingual support, open weights for self-hosting
ConsHeavy compute for local inference, not as cheap as smaller variants
ModelQwen/Qwen2.5-72B-Instruct
Scalelarge
BenchmarkingSource: Qwen2.5 technical report README (instruction-tuned).
- Qwen2.5-32B-Instruct (Alibaba / Qwen)Alibaba QwenGenerative
Released19 Sep 2024
Best forStrong quality with lower cost/latency than 72B
PriceOpen weights / provider-dependent API pricing
DeploymentLocal or API
ModalitiesText-only
Parameters32B instruct model, long-context variants available
ProsGood quality/performance balance
ConsStill requires substantial resources for local serving
ModelQwen/Qwen2.5-32B-Instruct
Scalemedium
BenchmarkingSource: Qwen2.5 README family eval table.
- Qwen2.5-14B-Instruct (Alibaba / Qwen)Alibaba QwenGenerative
Released19 Sep 2024
Best forMid-size production assistants, cost-aware coding/chat
PriceOpen weights / provider-dependent API pricing
DeploymentLocal or API
ModalitiesText-only
Parameters14B instruct model
ProsMuch easier to serve than 32B/72B, solid instruction following
ConsLower reasoning depth than larger Qwen models
ModelQwen/Qwen2.5-14B-Instruct
Scalemedium
BenchmarkingSource: Qwen2.5 README family eval table.
- Qwen2.5-7B-Instruct (Alibaba / Qwen)Alibaba QwenGenerative
Released19 Sep 2024
Best forLightweight assistants, edge/server cost-sensitive workloads
PriceOpen weights / provider-dependent API pricing
DeploymentLocal or API
ModalitiesText-only
Parameters7B instruct model
ProsFast, cheaper inference, widely deployable
ConsLower quality on harder reasoning/coding tasks
ModelQwen/Qwen2.5-7B-Instruct
Scaletiny
BenchmarkingSource: Qwen2.5 README family eval table.
- Qwen2.5-Coder-32B-Instruct (Alibaba / Qwen)Alibaba QwenGenerative (code-focused)
Released12 Nov 2024
Best forCode generation, refactoring, repo Q&A, code reasoning
PriceOpen weights / provider-dependent API pricing
DeploymentLocal or API
ModalitiesText-only
Parameters32B code-specialized instruct model
ProsStrong coding quality vs general-only models
ConsLarge model serving cost; for simple tasks smaller coder variants may be enough
ModelQwen/Qwen2.5-Coder-32B-Instruct
Scalelarge
BenchmarkingSource: Qwen2.5-Coder model card README (coding suites).
- Qwen2.5-VL-72B-Instruct (Alibaba / Qwen)Alibaba QwenGenerative
Released28 Jan 2025
Best forVision + text understanding, document/image question answering
PriceOpen weights / provider-dependent API pricing
DeploymentLocal or API
ModalitiesText + image
ParametersMultimodal (text + image) model family
ProsStrong multimodal capability in Qwen ecosystem
ConsHigher compute and serving complexity than text-only models
ModelQwen/Qwen2.5-VL-72B-Instruct
Scalelarge
BenchmarkingSource: Qwen2.5-VL technical README (vision QA / chart / doc benchmarks).
- Qwen3-Coder-Next (Alibaba Qwen)Alibaba QwenGenerative (code-focused)
ReleasedFeb 2026 (tech report 3 Mar 2026)
Best forLocal coding agents on modest hardware — repo-level coding, SWE tasks, tool use in Claude Code / Cline-style agents, fast non-thinking code completion. Unique: only 3B active of 80B — hybrid Gated DeltaNet (linear attention) + gated attention MoE that Qwen says matches models with 10–20× more active params; very fast and cheap to run locally.
PriceWeights free (Apache 2.0); infra cost only; hosted by third parties (e.g. Lambda)
DeploymentLocal / self-host (HF); SGLang, vLLM; llama.cpp / Ollama (GGUF); third-party inference APIs
ModalitiesText-only
ParametersMoE 80B total / 3B active; 48 layers in a hybrid layout 12 × (3 × Gated DeltaNet → MoE, 1 × Gated Attention → MoE); 512 experts, 10 active + 1 shared; hidden 2,048; 262,144-token context; non-thinking mode only (no <think> blocks); recommended temperature=1.0, top_p=0.95, top_k=40
ProsApache 2.0 open weights; tiny active compute (3B) — fast local inference; long 256K context; strong for its cost (SWE-bench Verified 70.6%); Base, Instruct, FP8 and GGUF builds; SGLang (≥0.5.8) / vLLM (≥0.15.0), OpenAI-compatible
ConsNo thinking mode — weaker on hard reasoning; well behind 2026 leaders on SWE-bench Pro (44.3%) and Terminal-Bench 2.0 (36.2%); still ~80B weights to load; text-only
ModelQwen/Qwen3-Coder-Next (also -Base, -FP8, GGUF)
Scalemedium
BenchmarkingSWE-bench Verified 70.6%; SWE-bench Pro 44.3%; Terminal-Bench 2.0 36.2%
- Qwen3.6-27B (Alibaba Qwen)Alibaba QwenReasoning / thinker
Released22 Apr 2026
Best forFlagship-level coding and agents on one GPU — SWE tasks, terminal agents, reasoning, multimodal (image / video) understanding; strong default open model for self-hosting. Unique: dense 27B that beats Qwen's own 397B MoE (Qwen3.5-397B-A17B) on SWE-bench Verified, SWE-bench Pro, Terminal-Bench 2.0 and SkillsBench; "thinking preservation" keeps earlier reasoning across turns for iterative coding.
PriceWeights free (Apache 2.0); infra cost only; Alibaba Cloud Model Studio / third-party APIs if hosted
DeploymentLocal / self-host (HF); SGLang, vLLM, KTransformers, Transformers; llama.cpp / Ollama (GGUF)
ModalitiesText + image + video in, text out
ParametersDense 27B; 64 layers in a hybrid layout 16 × (3 × Gated DeltaNet → FFN, 1 × Gated Attention → FFN); 262,144-token context, extendable to ~1,010,000 with YaRN; thinking mode on by default (non-thinking available); thinking preservation across turns; text + image + video input; recommended coding temperature=0.6, top_p=0.95, top_k=20
ProsApache 2.0 open weights; beats a 397B model on coding (SWE-bench Verified 77.2 vs 76.2; Terminal-Bench 2.0 59.3 vs 52.5); GPQA Diamond 87.8, AIME 2026 94.1; multimodal input; 1M context with YaRN; fits single high-memory GPU; 7M+ downloads, many GGUF builds; SGLang, vLLM, KTransformers, Transformers
ConsDense 27B is slower per token than small-active MoEs (e.g. Qwen3-Coder-Next); behind 2026 frontier models on SWE-bench Pro (53.5); superseded by Qwen3.8-27B (same architecture, ~4 months later); vendor-reported benches
ModelQwen/Qwen3.6-27B
Scalemedium
BenchmarkingSWE-bench Verified 77.2; SWE-bench Pro 53.5; Terminal-Bench 2.0 59.3; SkillsBench Avg5 48.2; GPQA Diamond 87.8; AIME 2026 94.1
- DeepSeek-V4.1-Flash (DeepSeek)DeepSeekReasoning / thinker
Released10 Sep 2026
Best forDefault DeepSeek model — agentic coding (SWE, terminal agents), reasoning, long-context work, image understanding; very cheap high-volume API or self-hosted open weights. Unique: new causal encoder-decoder MoE — 552B total but only 8B active while reading the prompt and 16B while generating; beats the much larger V4-Pro and matches Claude Opus 5 on Terminal-Bench 2.1 and DeepSWE at a fraction of the price.
PriceWeights free (MIT); API off-peak ~$0.15 / 1M input, ~$0.60 / 1M output; peak ~$0.30 / ~$1.20
DeploymentLocal / self-host (HF); DeepSeek API; DeepSeek web + mobile apps; NVIDIA NIM (build.nvidia.com)
ModalitiesText + image in, text out
ParametersMoE 552B total / 8B active (prefill) / 16B active (decode); causal encoder-decoder design; KV cache in 4-bit float (~890 bytes / token); 1M context; 384K max output; continuously controllable reasoning effort; image input (max 1,024 tokens per image)
ProsMIT open weights; frontier agentic coding (Terminal-Bench 2.1 90.6 vs Opus 5 89.1; DeepSWE v1.1 74.2%); very cheap ($0.15 / $0.60 off-peak); vision input; 1M context; small active compute
ConsStill ~552B weights to self-host; vendor-reported benches; image input limited to 1,024 tokens per image; peak-hour prices are double
Modeldeepseek-ai/DeepSeek-V4.1-Flash (API: deepseek-v4-flash now routes to V4.1 Flash)
Scalelarge / frontier
BenchmarkingTerminal-Bench 2.1 90.6 (Claude Opus 5 89.1); DeepSWE v1.1 74.2% (Opus 5 74.0%, V4-Pro 62.7%)
- DeepSeek-V4-Pro (DeepSeek)DeepSeekReasoning / thinker
Released24 Apr 2026 (preview); 0813 update Aug 2026
Best forLarge text-only DeepSeek model for coding, reasoning, browsing agents and long-context work; teams pinned to V4-Pro output style. Unique: 1.6T MoE with hybrid Compressed Sparse Attention + Heavily Compressed Attention for 1M context, Manifold-Constrained Hyper-Connections and Muon optimizer; DSpark speculative-decoding variant available.
PriceWeights free (MIT); API off-peak ~$0.66 / 1M input; peak ~$1.32 / 1M input, ~$3.96 / 1M output
DeploymentLocal / self-host (HF); DeepSeek API; vLLM, SGLang, Docker Model Runner; HuggingChat
ModalitiesText-only
ParametersMoE 1.6T total / 49B active; hybrid attention (Compressed Sparse Attention + Heavily Compressed Attention); Manifold-Constrained Hyper-Connections; Muon optimizer; 1M context; 384K max output; text-only; DSpark drafter checkpoint (DeepSeek-V4-Pro-DSpark)
ProsMIT open weights; strong benches (SWE-bench Verified 80.6%, GPQA Diamond 90.1%, LiveCodeBench 93.5%, BrowseComp 83.4%); 1M context; API kept alive after Sep 14 due to user demand
ConsBeaten by the cheaper V4.1-Flash on agentic coding; text-only; ~4× V4.1-Flash price; huge to self-host; no V4.1-Pro yet
Modeldeepseek-ai/DeepSeek-V4-Pro (API: deepseek-v4-pro)
Scalelarge / frontier
BenchmarkingSWE-bench Verified 80.6%; SWE-bench Pro 55.4%; Terminal-Bench 2.0 67.9%; GPQA Diamond 90.1%; LiveCodeBench 93.5%; HLE (with tools) 48.2%; BrowseComp 83.4%
- Kimi K3 (Moonshot AI)Moonshot AI (Kimi)Reasoning / thinker
Released16 Jul 2026 (open weights 27 Jul 2026)
Best forFrontier coding and long-horizon agents (SWE, terminal, browsing / deep research, computer use), 1M-context repo and document work, professional knowledge tasks, vision + video understanding. Unique: largest open-weights model at launch (2.8T MoE) — native MXFP4 checkpoint from quantization-aware training, Kimi Delta Attention (KDA) hybrid for 1M context, and #3 on GDPval-AA v2 behind only Claude Fable 5 Max and GPT-5.6 Sol Max.
PriceWeights free (Kimi K3 License); API ~$3 / 1M input, ~$15 / 1M output (flat at any context length)
DeploymentLocal / self-host (HF); Kimi API (platform.kimi.ai); OpenRouter; vLLM, SGLang, TokenSpeed
ModalitiesText + image + video
ParametersMoE 2.8T total / 104B active; 896 routed experts, 16 active per token + 2 shared; 93 layers (69 Kimi Delta Attention + 24 Gated MLA, 1 dense); hidden 7,168, 96 heads; Attention Residuals (AttnRes); 1,048,576-token context; vision encoder MoonViT-V2 (401M); MXFP4 weights / MXFP8 activations (QAT); thinking always on, reasoning effort low / high / max (default max); launched 16 Jul 2026, open weights 27 Jul 2026
ProsOpen weights; frontier-level agent / coding benches (Terminal-Bench 2.1 within 0.5 pt of GPT-5.6 Sol, BrowseComp above Claude Fable 5); flat $3 / $15 pricing at any context length; native vision + video input; 1M context; native MXFP4 checkpoint lowers serving memory vs BF16
ConsEnormous to self-host (2.8T weights — multi-node even at MXFP4); custom Kimi K3 License (not MIT / Apache — check commercial terms); thinking cannot be turned off; still behind GPT-5.6 Sol on GPQA and DeepSWE; mostly vendor-reported numbers; pricier than K2.x and MiMo-V2.6
Modelmoonshotai/Kimi-K3 (API: kimi-k3; OpenRouter: moonshotai/kimi-k3)
Scalelarge / frontier
Benchmarking- FrontierSWE81.2%
- Terminal-Bench 2.188.3%
- DeepSWE67.5%
- GPQA-Diamond93.5%
- BrowseComp91.2%
- HLE-Full (w/ tools)56.0
Source: Moonshot-reported (huggingface.co/moonshotai/Kimi-K3 card, github.com/MoonshotAI/Kimi-K3). Also: HLE-Full 43.5 (no tools); MMMU-Pro 81.6 / 83.4; OSWorld 2.0 58.3; GDPval-AA v2 Elo 1,687 (#3); Artificial Analysis Intelligence Index v4.3: 44 (third-party, checked 7 Sep 2026).
- moonshot-v1-128k-vision-previewMoonshot AI (Kimi)Generative
Released15 Jan 2025
Best forHeavy multimodal context (long system + user + images) in one shot
PriceToken-based; see Moonshot pricing docs
DeploymentAPI
ModalitiesText + image
ParametersVision model id for ~128K context tier (preview)
ProsLargest v1 vision context tier in the name
ConsPreview; most expensive/heaviest when you use full context
Modelmoonshot-v1-128k-vision-preview
Scalelarge
BenchmarkingSource: BenchLM (Moonshot v1 family profile; not tier-specific to 128k preview id).
- moonshot-v1-32k-vision-previewMoonshot AI (Kimi)Generative
Released15 Jan 2025
Best forLonger multimodal chats / more image+text context in one request
PriceToken-based (vision token accounting); see Moonshot pricing docs
DeploymentAPI
ModalitiesText + image
ParametersVision model id for ~32K context tier (preview); same image rules as other Moonshot vision models
ProsMore room for instructions + image context than 8K tier
ConsPreview; higher token use and cost than 8K when you fill context
Modelmoonshot-v1-32k-vision-preview
Scalemedium
BenchmarkingSource: BenchLM (Moonshot v1 family profile; not tier-specific to 32k preview id).
- moonshot-v1-8k-vision-previewMoonshot AI (Kimi)Generative
Released15 Jan 2025
Best forImage + text in one request; short prompts and small multimodal turns
PriceToken-based chat pricing (vision uses dynamic image/video tokens); see Moonshot pricing docs
DeploymentAPI
ModalitiesText + image
ParametersVision model id for ~8K context tier (preview); images: png, jpeg, webp, gif; see Moonshot vision guide for request format (base64 / file id)
ProsLower context cost vs larger tiers when input fits in 8K
ConsPreview name may change; long images + long text can hit limits faster
Modelmoonshot-v1-8k-vision-preview
Scaletiny
BenchmarkingSource: BenchLM (Moonshot v1 family profile; not tier-specific to 8k preview id).
- MiniMax M3 (MiniMax)MiniMax AIReasoning / thinker
Released1 Jun 2026
Best forFrontier coding + long-horizon agents, repo-scale / 1M-context work, office/search workflows, native image+video understanding, desktop/computer-use style agents. Unique: open-weight model that packs frontier coding/agentic strength, 1M context via MiniMax Sparse Attention (MSA), and native multimodality (mixed-modality training from step 0) in one stack.
PriceOpen weights (MiniMax Community License); API / Token Plan billed — see https://platform.minimax.io pricing
DeploymentLocal / self-host (HF); MiniMax API; MiniMax Code / Agent products; SGLang, vLLM, Transformers, KTransformers, unsloth, ATOM
ModalitiesText + image + video
ParametersMoE ~428B total / ~23B active; up to 1M context; attention = MiniMax Sparse Attention (MSA) on GQA — Index branch picks top KV blocks per group, Main branch runs exact softmax only on those blocks (not a separate model id); thinking modes via API thinking: enabled / adaptive / disabled; recommended decode temperature=1.0, top_p=0.95
ProsOpen weights on HF; 1M context with large MSA speedups vs dense attention (~9× prefill / ~15× decode vs M2 at 1M per MiniMax; MSA paper ~28.4× attention FLOPs vs GQA at 1M); strong coding/agent benches; native text+image+video; shippable via SGLang / vLLM / Transformers / etc.; API + Token Plan + MiniMax Code
ConsHuge to self-host; MiniMax Community License (not Apache/MIT — commercial use needs attribution / notice or written auth above revenue thresholds); MSA is architecture detail, not a catalog Scale row; vendor benches dominate public numbers
ModelMiniMaxAI/MiniMax-M3
Scalelarge / frontier
BenchmarkingSource: MiniMax-reported (minimax.io/blog/minimax-m3; Hugging Face MiniMaxAI/MiniMax-M3). Not a third-party rerun. MSA: arxiv.org/abs/2606.13392.
- MiniMax M2.7 (MiniMax)MiniMax AIReasoning / thinker
Released18 Mar 2026
Best forAgents, coding, software engineering, office workflows, long-context reasoning; flagship text line
PriceSee https://platform.minimax.io pricing (token plans / pay-as-you-go)
DeploymentAPI (OpenAI- or Anthropic-compatible SDKs per docs)
ModalitiesText-only
ParametersMoE-class stack (~230B total / ~100B active per public materials); very long context (~204.8K-class); tools / Anthropic-compatible API path
ProsStrong real-world engineering and agentic positioning; highspeed variant for latency-sensitive paths
ConsAPI-only (no open weights); pricing/quotas region- and account-dependent
ModelMiniMax-M2.7
Scalelarge / frontier
BenchmarkingSource: MiniMax M2.7 model page (minimax.io/models/text/m27).
- MiniMax M2.7-highspeed (MiniMax)MiniMax AIReasoning / thinker
Released18 Mar 2026
Best forSame tasks as M2.7 when you need lower latency / higher throughput
PriceSee MiniMax pricing (often differs from base M2.7)
DeploymentAPI
ModalitiesText-only
ParametersSame capability tier as M2.7; faster inference (vendor-tuned routing)
ProsSignificantly faster than base M2.7 for similar quality class
ConsAPI-only; may differ slightly in throughput vs base under load
ModelMiniMax-M2.7-highspeed
Scalelarge / frontier
BenchmarkingMiniMax docs: same eval results as M2.7, higher throughput path.
- MiniMax M2.5 (MiniMax)MiniMax AIGenerative
Released12 Feb 2026
Best forCode generation, refactoring, polyglot coding, strong value tier before M2.7
PriceSee MiniMax pricing
DeploymentAPI
ModalitiesText-only
ParametersLong context (~204.8K-class per docs); code-optimized positioning
ProsPeak value tier in the M2 text line for coding-heavy workloads
ConsSuperseded for absolute frontier by M2.7 on vendor charts; API-only
ModelMiniMax-M2.5
Scalelarge
BenchmarkingThird-party SWE-bench Verified aggregates (model family tier; not SWE-Pro).
- MiniMax M2.5-highspeed (MiniMax)MiniMax AIGenerative
Released12 Feb 2026
Best forSame as M2.5 with lower latency
PriceSee MiniMax pricing
DeploymentAPI
ModalitiesText-only
ParametersSame performance class as M2.5; faster inference
ProsFast M2.5-class option for high-volume coding
ConsAPI-only
ModelMiniMax-M2.5-highspeed
Scalelarge
BenchmarkingSame published tier as M2.5; vendor lists highspeed for latency.
- MiniMax M2.1 (MiniMax)MiniMax AIGenerative
Released23 Dec 2025
Best forCode, reasoning, refactoring; legacy M2 line still listed for stable integrations
PriceSee MiniMax pricing
DeploymentAPI
ModalitiesText-only
ParametersMoE-style (~230B total, ~10B activated per token per docs); code-focused
ProsMature tier; often cheaper than newest flagship
ConsLegacy relative to M2.5 / M2.7; API-only
ModelMiniMax-M2.1
Scalelarge
BenchmarkingSource: MiniMax-AI/MiniMax-M2.1 GitHub README benchmark table.
- MiniMax M2 (MiniMax)MiniMax AIReasoning / thinker
Released27 Oct 2025
Best forLong output and agentic text (function calling, streaming) on older M2 generation
PriceSee MiniMax pricing
DeploymentAPI
ModalitiesText-only
Parameters~200K context; up to ~128K output (incl. chain-style content per docs)
ProsEstablished M2 generation; long outputs
ConsLegacy vs M2.5 / M2.7; API-only
ModelMiniMax-M2
Scalelarge
BenchmarkingSource: MiniMax-AI/MiniMax-M2.1 GitHub README (M2 column).
- M2-her (MiniMax)MiniMax AIGenerative
Released23 Jan 2026
Best forRoleplay, multi-character dialogue, long-horizon character interaction
PriceSee MiniMax pricing
DeploymentAPI
ModalitiesText-only
ParametersText chat tuned for character and emotional expression
ProsSpecialized for interactive fiction / persona use cases
ConsNot a general coding frontier model; API-only
ModelM2-her
ScaleN/A
BenchmarkingRoleplay-specialized SKU — not leaderboard-optimized coding evals.
- GLM-5.2 (Z.ai / Zhipu)Z.ai (Zhipu)Reasoning / thinker
Released13 Jun 2026 (Coding Plan); API + open weights ~16 Jun 2026
Best forLong-horizon coding and agent tasks (hours-scale projects, SWE, terminal, MCP tools), 1M-context repo/agent workflows, reasoning-heavy STEM. Unique: flagship open GLM after 5.1 that first pairs that long-horizon strength with a solid 1M-token window; IndexShare sparse attention (shared indexer across layers) + stronger MTP for throughput — MIT weights, no regional lock.
PriceWeights free (MIT); Z.ai / third-party APIs billed if used hosted
DeploymentLocal / self-host (HF, ModelScope); Z.ai API; vLLM, SGLang, Transformers, xLLM, ktransformers; NIM / FP8 paths where published
ModalitiesText-only
ParametersMoE ~753B total / ~39–40B active (public cards also cite ~743B class); native context up to ~1,048,576 tokens (to 128K output class per docs); IndexShare reuses one lightweight indexer across sparse-attn layers (cuts indexer FLOPs ~2.9× at 1M per Z.ai — architecture feature, not a separate model id); MTP extended (draft tokens up vs GLM-5/5.1); thinking-mode behavior retained; checkpoints BF16, native FP8, AMD Quark MXFP4; released mid-June 2026
ProsMIT open weights (HF + ModelScope); strong vendor coding/agent benches vs prior GLM and competitive with closed flagships on several long-horizon tables; real 1M context path; vLLM / SGLang / Transformers / xLLM / ktransformers; FP8 practical for single 8×H200-class nodes
ConsHuge to self-host (FP8 ~hundreds of GB; BF16 multi-node); vendor-reported benches dominate; IndexShare is architecture detail, not a catalog Scale row; text-first (not a native multimodal pack like MiniMax M3)
Modelzai-org/GLM-5.2 (also zai-org/GLM-5.2-FP8)
Scalelarge / frontier
Benchmarking- SWE-Bench Pro62.1%
- Terminal-Bench 2.1 (Terminus-2)81.0%
- Terminal-Bench 2.1 (best harness)82.7%
- FrontierSWE (Dominance)74.4%
- MCP-Atlas (Public)76.8%
- GPQA-Diamond91.2%
- HLE (text)40.5%
- HLE (w/ Tools)54.7%
- AIME 202699.2%
- Tool-Decathlon48.2%
- PostTrainBench34.3%
- SWE-Marathon13.0%
Source: Z.ai-reported (huggingface.co/blog/zai-org/glm-52-blog; zai-org/GLM-5.2 card). Not a third-party rerun. IndexShare: arxiv.org/abs/2603.12201.
- JEV (TypeSafe AI)TypeSafe AIClassification / decision
Released15 Sep 2026 (early access)
Best forFast, cheap, typed decisions inside programs — routing, classification, scoring, yes/no gates, agent step choices, moderation / triage verdicts. Unique: a "decision engine" rather than a text generator — one forward pass returns a schema-constrained value (Choice / Score / Bool), so output can't be invalid and needs no parsing; vendor claims ~20–200× faster and ~40–400× cheaper than frontier LLMs on decision tasks.
Price~$0.042 per 1M input tokens; output tokens free
DeploymentAPI (TypeSafe AI, waitlist)
ModalitiesText-only
ParametersParameter count / architecture not published; trained with RLCD (Reinforcement Learning for Calibrated Decisions) instead of RLHF; "parallel sampling" — all outputs returned in a single query, no autoregressive token-by-token generation; output limited to a predefined schema (Choice / Score / Bool types); latency ~70–500 ms per vendor (independent measurement ~236–276 ms) vs ~3–30 s for an LLM. Inferred (not confirmed): closer to a modern classifier head (BERT-style) over rich program-state context than to a GPT-style generator
ProsVery low latency and cost (output tokens free); always returns a valid typed value (no JSON repair / parsing); calibrated scores via RLCD; drop-in for decision points in agents and pipelines where an LLM is overkill
ConsClosed / API-only, waitlist access; no open weights or self-hosting; no free-text output (can't explain, write, or chat); architecture, size and training data undisclosed; benchmarks mostly vendor-reported
ModelJEV (TypeSafe AI API; exact model id: confirm in TypeSafe console)
ScaleN/A
BenchmarkingSource: vendor claims (20–200× faster, 40–400× cheaper vs LLMs) plus the Laya model card (huggingface.co/convaiinnovations/laya), which reports Jev 0.727 on its decision benchmark. No public standard-suite scores. Open-weights alternative: Laya (Convai Innovations, 421M, ModernBERT-based).
Alternatives- LayaConvai Innovations · open weights · 421M · ModernBERT-based
- KEVdetails TBD
- nimbledetails TBD
- jevlikedetails TBD
- mini jevdetails TBD
- jeffdetails TBD
- semifdetails TBD
- MiMo-V2.6-Pro (Xiaomi)Xiaomi MiMoReasoning / thinker
Released21 Sep 2026
Best forFrontier-class coding and long-horizon agents (SWE, terminal, tool use), 1M-context repo / multi-session agent runs, omnimodal understanding (text + image + video + audio), cybersecurity tasks. Unique: top open-weights model on the Artificial Analysis Intelligence Index at launch (~46, tied with Grok 4.7, ahead of DeepSeek V4.1) — a ~1T MIT-licensed omnimodal MoE priced around 1/10 of closed flagships.
PriceWeights free (MIT); API ~$0.435 / 1M input, ~$0.004 / 1M cached input, ~$0.87 / 1M output
DeploymentLocal / self-host (HF); Xiaomi MiMo API; OpenRouter; SGLang (tp=16, dp=2), vLLM (tp=8, multi-node)
ModalitiesText + image + video + audio
ParametersSparse MoE 1.02T total / 42B active; 70 layers (60 sliding-window SWA + 10 global attention, 128-token window); hidden 6144; 128 query / 8 KV heads; 384 routed experts, 8 active; 1M context; MTP via 5-layer DFlash-style speculative drafter; vision encoder 681M, audio 308M AudioTokenizer + 127M patch encoder; scaled RL post-training (~3.5–3.7B tokens per RL step); released 22 Sep 2026
ProsMIT open weights; native omnimodal in one model; 1M context; very cheap API vs Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30); strong agent / coding benches; SGLang + vLLM support
ConsHuge to self-host (multi-node, e.g. SGLang tp=16); not an across-the-board win vs closed models (Claude Opus 5 still ahead on several benches incl. DeepSWE); most agent evals vendor-run, not independently verified
ModelXiaomiMiMo/MiMo-V2.6-Pro-RL (API: mimo-v2.6-pro)
Scalelarge / frontier
Benchmarking- AA Intelligence Index~46
- DeepSWE v1.171.9%
- Terminal-Bench 2.189.9%
- Toolathlon-Verified76.9%
- AutomationBench53.1%
- CyberGym94.0%
Source: Xiaomi-reported (huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL card; launch post mimo.mi.com/docs/en-US/news/latest/v2-6). Launch post lists DeepSWE v1.1 as 72.6 (up from 58.4 in V2.5); card lists 71.9. Not a third-party rerun.
- MiMo-V2.6-Flash (Xiaomi)Xiaomi MiMoReasoning / thinker
Released21 Sep 2026
Best forCheap, fast coding / agent workloads, high-volume omnimodal tasks, 1M-context work on a budget, self-hosting on a single node. Unique: ~15B-active omnimodal MoE that stays close to Pro on agent benches (and beats it on CyberGym) at ~1/3 of Pro's price.
PriceWeights free (MIT); API ~$0.14 / 1M input, ~$0 cached input, ~$0.28 / 1M output
DeploymentLocal / self-host (HF); Xiaomi MiMo API; OpenRouter; SGLang (tp=8), vLLM (tp=4)
ModalitiesText + image + video + audio
ParametersSparse MoE 309B total / 15B active; 48 layers (39 SWA + 9 global attention); hidden 4096; 256 routed experts, 8 active; 1M context; MTP = 5-layer SWA DFlash-style drafter predicting 7 tokens per forward pass; same vision (681M) and audio (308M + 127M) encoders as Pro; released 22 Sep 2026
ProsMIT open weights; very cheap API; single-node self-host (SGLang tp=8 / vLLM tp=4); native omnimodal; 1M context; fast decode via MTP drafting
ConsBelow Pro on hard coding / reasoning; still ~300B weights to host; vendor-reported benches
ModelXiaomiMiMo/MiMo-V2.6-Flash-RL (API: mimo-v2.6-flash)
Scalemedium
BenchmarkingSource: Xiaomi-reported (huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL card). Launch post lists DeepSWE v1.1 as 65.7 (up from 48.8 in V2.5); card lists 67.9. Not a third-party rerun.
- MiMo-V2.6-Pro-UltraSpeed (Xiaomi)Xiaomi MiMoReasoning / thinker
Released21 Sep 2026
Best forLatency-critical use of Pro-level quality — interactive coding agents, real-time assistants, long generations where wall-clock time matters more than cost.
Price~$4.35 / 1M input, ~$8.70 / 1M output (10× Pro)
DeploymentAPI (Xiaomi MiMo API, OpenRouter)
ModalitiesText + image + video + audio
ParametersSpeed-optimized serving tier of MiMo-V2.6-Pro (no separate weights published); Xiaomi claims up to 20× Pro output speed at the same quality (OpenRouter says ~10×); serving mechanism not disclosed; 1M context; released 21 Sep 2026
ProsPro quality at much higher output speed (per vendor); same API surface as Pro
Cons10× Pro's price; API-only tier; speed claims not yet independently confirmed
Modelmimo-v2.6-pro-ultraspeed (OpenRouter: xiaomi/mimo-v2.6-pro-ultraspeed)
Scalelarge / frontier
BenchmarkingSame quality as MiMo-V2.6-Pro per Xiaomi; speed is vendor-claimed ("up to 20×"), not yet independently measured — early OpenRouter stats ~125 tok/s vs Pro ~130 tok/s.
- Claude Fable 5.1 (Anthropic)AnthropicReasoning / thinker
Released1 Sep 2026
Best forHardest open-ended work — frontier research, long-horizon agents, complex architecture and reasoning, high-stakes analysis; top generally available Claude. Unique: public, guard-railed version of the Mythos-class model — requests classified high-risk (cybersecurity, biology) are downgraded to Opus 4.8; ~25% cheaper than Fable 5 per typical task (up to ~45% on agentic workloads).
Price~$10 / 1M input, ~$50 / 1M output; cache reads 2.5% of input; Batch 50% off
DeploymentClaude API, Amazon Bedrock, Google Cloud Vertex AI; Claude apps
ModalitiesText + image in, text out
Parameters1M context; 128K max output; adaptive thinking; text + image input; knowledge cutoff Jun 2026
ProsMost capable generally available Claude; very cheap cache reads (2.5% of input price); strong across coding, research, computer use (OSWorld 2.0 80.7%, HLE with tools 65.6%)
ConsMost expensive Claude ($10 / $50); Opus 5.5 now matches or beats it on most benches at 40% of the price; high-risk requests silently routed to Opus 4.8
Modelclaude-fable-5-1 (Bedrock: anthropic.claude-fable-5-1)
Scalelarge / frontier
BenchmarkingTerminal-Bench 4.0 55.8%; FrontierCode v1.1 50.3%; CursorBench 4.0 51.8%; HLE (with tools) 65.6%; OSWorld 2.0 (partial) 80.7%; AutomationBench 31.4%
- Claude Opus 5.5 (Anthropic)AnthropicReasoning / thinker
Released22 Sep 2026
Best forHard coding and agent work at scale — SWE, terminal agents, computer use, knowledge work; best price / quality point for Fable-class tasks. Unique: leads Fable 5.1 on SWE-bench Pro, Terminal-Bench 4.0, CursorBench, HLE and OSWorld 2.0 at 40% of Fable's per-token price; highest SWE-bench Pro on record (89.9%).
Price~$4 / 1M input, ~$20 / 1M output; cache reads ~$0.20 / 1M; Batch 50% off
DeploymentClaude API, Amazon Bedrock, Google Cloud Vertex AI; Claude apps, Claude Code
ModalitiesText + image in, text out
Parameters1M context; 128K max output (Batch: 300K with beta header); adaptive thinking always on, default medium effort (up to xhigh / max); text + image input; knowledge cutoff Jun 2026
ProsTop coding benches (SWE-bench Pro 89.9%, Terminal-Bench 4.0 66.4% — ~10 pt above Fable 5.1 and GPT-6 Astra); Fable-level work at $4 / $20; cheap cache reads ($0.20 / 1M)
ConsPricier than Sonnet 5.5 for routine work; thinking cannot be fully disabled; all numbers vendor-reported
Modelclaude-opus-5-5 (Bedrock: anthropic.claude-opus-5-5)
Scalelarge / frontier
BenchmarkingSWE-bench Pro 89.9%; Terminal-Bench 4.0 (xhigh) 66.4%; DeepSWE v1.1 74.2%; CursorBench 4.0 57.8%; FrontierCode v1.1 54.4%; HLE (with tools) 67.7%; OSWorld 2.0 (partial) 81.8% / (strict) 48.7%; AutomationBench 40.0%; GDPval-AA v2.1 Elo 1846
- Claude Sonnet 5.5 (Anthropic)AnthropicReasoning / thinker
Released28 Sep 2026
Best forDefault production model — chat, coding, agents, computer use, vision; best speed / cost balance. Unique: close to Opus 5.5 on key benches at half its price; ~30% faster and up to ~30% cheaper per task than Sonnet 5.
Price~$2 / 1M input, ~$10 / 1M output; cache reads 10% of input; Batch 50% off
DeploymentClaude API, Amazon Bedrock, Google Cloud Vertex AI; Claude apps, Claude Code
ModalitiesText + image in, text out
Parameters1M context; 128K max output; adaptive thinking; text + image input; knowledge cutoff Jun 2026
ProsStrong coding (SWE-bench Pro 81.3%) and computer use (OSWorld 2.1 80.1%) at $2 / $10; fast; recommended default for most builders
ConsBelow Opus 5.5 / Fable 5.1 on the hardest tasks; few full benchmark tables published yet
Modelclaude-sonnet-5-5 (Bedrock: anthropic.claude-sonnet-5-5)
Scalelarge
BenchmarkingSWE-bench Pro 81.3%; OSWorld 2.1 80.1%
- Claude Haiku 4.5 (Anthropic)AnthropicGenerative
Released15 Oct 2025
Best forLow-latency chat, high-volume routing, sub-agents, cost-sensitive workloads
Price~$1 / 1M input, ~$5 / 1M output; cache reads 10% of input; Batch 50% off
DeploymentClaude API, Amazon Bedrock, Google Cloud Vertex AI
ModalitiesText + image in, text out
Parameters200K context; 64K max output; extended thinking; knowledge cutoff Feb 2025; still the current Haiku (Haiku 5.5 announced "in the coming weeks")
ProsFastest and cheapest current Claude
ConsOlder generation than the 5.x line; smaller context (200K vs 1M); successor Haiku 5.5 coming soon
Modelclaude-haiku-4-5-20251001 (alias: claude-haiku-4-5)
Scalemedium
BenchmarkingSWE-bench Verified (public aggregate) 73.3%; WebArena 53.1% (Anthropic Claude model family WebArena table, single-policy)
- Claude Mythos 5.1 (Anthropic)AnthropicReasoning / thinker (restricted)
Released1 Sep 2026
Best forRestricted high-capability work — defensive cybersecurity, vulnerability research, security analysis — for vetted organizations only. Unique: the less-restricted Mythos-class model behind Fable; earlier Mythos Preview found 271 Firefox vulnerabilities for Mozilla.
PriceNot public
DeploymentTrusted-access program only
ModalitiesText + image in, text out
ParametersMythos-class frontier model; details not published publicly
ProsFewest guardrails of any Claude for approved security research
ConsNot generally available — limited trusted-access program only (150+ organizations, 15+ countries); no public pricing or benchmarks; access briefly revoked in Jun 2026 over US export directives
ModelN/A (trusted-access program)
Scalelarge / frontier
Benchmarking- BenchmarkingN/A
Not published
- CONTEXT-1ChromaAgentic retrieval
Released26 Mar 2026
Best forMulti-hop retrieval, agentic search paired with a frontier reasoning model
PriceOpen weights (inference cost on your infra)
DeploymentHugging Face, local
ModalitiesText-only
Parameters~20B params; query decomposition, iterative corpus search, in-loop context editing
ProsBuilt for complex multi-hop retrieval as a sub-agent
ConsSpecialized workflow; not a general chat or coding model
Modelchromadb/context-1
ScaleN/A
BenchmarkingSource: Chroma Context-1 tech report — Final Answer Found (F1) on retrieval-led tasks.
- Groq Compound (Groq)GroqGenerative (orchestrated)
Released15 Apr 2025 (beta); GA 4 Sep 2025
Best forTool-orchestrated search and routing across Groq-hosted models (e.g. open-weight ~120B-class paths per Groq)
PriceSee Groq Cloud pricing
DeploymentAPI (Groq)
ModalitiesText-only
ParametersProduct routes across retrieval + generative models; not one fixed public parameter count
ProsLow-latency Groq inference; unified compound surface
ConsWhich sub-model runs can change; not a single static checkpoint
Modelgroq-compound
Scalelarge
BenchmarkingRouting product; no single static benchmark row.
- Tiny Recursive Model (TRM)Samsung (SAIL Montréal)Reasoning / thinker
Released6 Oct 2025 (paper arXiv 2510.04871)
Best forStructured puzzle reasoning (ARC-AGI, Sudoku, mazes), recursive-reasoning research
PriceFree (open source)
DeploymentLocal
ModalitiesStructured grids / puzzles (not free-form conversational text)
Parameters~7M parameters, iterative latent and answer refinement (paper: arXiv:2510.04871)
ProsOrders-of-magnitude smaller than LLMs on comparable reasoning tasks, MIT-licensed repo
ConsNot a general chat model, embedding, or reranker; no official hosted API; GPU stack for train/eval
ModelSamsungSAILMontreal/TinyRecursiveModels
Scaletiny
BenchmarkingARC grid tasks; not MMLU-style chat.
- Inkling (Thinking Machines Lab)Thinking Machines LabReasoning / thinker
Released15 Jul 2026
Best forOpen-weights multimodal base for fine-tuning / domain adaptation (agents, coding, chat, RAG, vision+audio→text); Tinker customization workflows. Unique: designed as a customizable base (not leaderboard-first) with controllable thinking effort, native text/image/audio reasoning (video in pretrain), strong calibration/epistemics focus, and first-class fine-tune path on Tinker — built as the reasoning backbone for their interaction-models line.
PriceWeights free (open); Tinker / third-party APIs billed (Inkling on Tinker had a limited-time 50% discount — check current Tinker docs)
DeploymentLocal / self-host (HF weights); Tinker (fine-tune + playground); APIs (Together, Fireworks, Modal, Databricks, Baseten, HF Inference Providers)
ModalitiesText + image + audio (+ video in pretrain; text output)
ParametersMoE transformer ~DeepSeek-V3-style (256 routed + 2 shared experts, 6 routed active/token); 975B total / 41B active; up to 1M context (Tinker offers 64K / 256K); pretrained ~45T tokens (text, image, audio, video); encoder-free vision (40×40 patches) + audio (dMel); relative PE + sliding/global attn (5:1); BF16 + NVFP4 checkpoints; family also has Inkling-Small preview (276B / 12B active)
ProsFull open weights (Apache 2.0); multimodal in / text out; effort dial for cost vs quality; strong open-weights audio/safety (FORTRESS); calibrated answers / forecasting-oriented training; ecosystem day-0 (HF, SGLang, vLLM, llama.cpp, Together, Fireworks, …)
ConsNot the strongest overall model (lab’s own claim); huge to self-host; output is text-only for now; Inkling-Small weights still preview/not fully released; Tinker context options shorter than max 1M
Modelthinkingmachines/Inkling (also thinkingmachines/Inkling-NVFP4)
Scalelarge / frontier
Benchmarking- HLE (text)29.7%
- HLE (with tools)46.0%
- AIME 202697.1%
- GPQA Diamond87.2%
- SWE-Bench Verified77.6%
- Terminal Bench 2.163.8%
- IFBench79.8%
- Global-MMLU-Lite88.7%
- MMMU Pro73.5%
- VoiceBench91.4%
- FORTRESS (adversarial)78.0%
- FORTRESS (benign)95.9%
Source: lab-reported at effort=0.99 (Thinking Machines / Hugging Face Inkling model card). Not a third-party rerun.
- Needle (Cactus Compute)Cactus ComputeFunction calling / tool router
Released12 May 2026
Best forOn-device / edge tool routing — single-shot function calling (user query + tool schemas → tool name + JSON args) on phones, watches, CPUs; cheap local dispatch step in MCP / tool-using agents. Unique: a 26M "tool router" (~14MB INT4) — Cactus argues tool calling is query → tool → fill JSON, not deep reasoning, so a tiny specialist can handle it locally instead of a cloud LLM.
PriceFree (open source, MIT)
DeploymentLocal / on-device (phone, watch, CPU); GitHub, Hugging Face
ModalitiesText-only (query + tool schemas in, JSON function call out)
Parameters26M params; distilled from Gemini 3.1 Flash Lite (teacher); Simple Attention Network (SAN) backbone — attention + gating, no MLP/FFN layers; ~14MB at INT4; claimed ~6,000 tok/s prefill / ~1,200 tok/s decode on consumer hardware
ProsTiny footprint, runs fully on-device (no cloud round-trip); very fast; MIT license; beats ~10× larger models (FunctionGemma-270M, Qwen-0.6B, etc.) on single-shot tool calling per Cactus
ConsNot a chat or reasoning model; weak on ambiguity, conversation, multi-step agent loops; single-shot only; vendor-reported numbers
Modelcactus-compute/needle (GitHub; HF repo id: confirm)
Scaletiny / distilled
BenchmarkingSource: Cactus Compute blog (vendor-reported). Single-shot tool calling above FunctionGemma-270M and Qwen-0.6B (~10× larger); weaker on conversation, ambiguity and multi-step agent loops. Not a third-party rerun.
- Ornith-1.0 (DeepReinforce)DeepReinforce (Ornith)Reasoning / thinker
Released25 Jun 2026
Best forAgentic coding — SWE tasks, terminal agents, repo-level fixes, MCP / tool use; self-hosted coding agents at several sizes (laptop 9B up to 397B server). Unique: "self-scaffolding" — the model builds its own agent harness (task plan, tool calls, checks on intermediate results) and then solves the task inside it; during RL the scaffold is learned together with the policy instead of being a hand-written harness.
PriceWeights free (MIT); infra cost only (no first-party API listed)
DeploymentLocal / self-host (HF); vLLM, SGLang, llama.cpp / Ollama (GGUF), MLX
ModalitiesText-only
ParametersFamily of 4 sizes — 9B dense, 31B dense, 35B MoE (~3B active), 397B MoE (flagship, figures below); post-trained on Gemma 4 and Qwen 3.5 bases (not trained from scratch); 262,144-token context (~1M via YaRN); text-only; self-scaffolding RL with async pipeline-RL, token-level GRPO + staleness weighting; released 25 Jun 2026 (successor Ornith-1.5 also out)
ProsMIT open weights, no regional limits; strong agentic coding for open models; 35B MoE scores 64.2 on Terminal-Bench 2.1 with ~3B active; OpenAI-compatible serving; GGUF / MLX paths for small sizes
ConsPost-train of Gemma 4 / Qwen 3.5, not a new base; text-only; behind Claude Opus 4.8 and GLM-5.2 on several tables; vendor-reported benches; already superseded by Ornith-1.5
Modelornith-ai/Ornith-1.0-397B (also -9B, -31B, -35B)
Scalelarge / frontier
BenchmarkingSource: DeepReinforce-reported (huggingface.co/ornith-ai/Ornith-1.0-397B card). Also: Terminal-Bench 2.1 with Claude Code harness 78.2%; SWE-bench Multilingual 78.9%; Claw-Eval avg 77.1%. 397B beats Claude Opus 4.7 on Terminal-Bench 2.1 (77.5 vs 70.3). Not a third-party rerun.
- LongCat-2.0 (Meituan)Meituan (LongCat)Reasoning / thinker
Released30 Jun 2026
Best forAgentic coding and repo-level software engineering (SWE, terminal agents, multilingual codebases), deep-research / browsing agents, 1M-context work; cheap API backend for Claude Code / Codex / OpenCode-style tools. Unique: first model at this scale trained from scratch AND served entirely on Chinese-made AI chips (50,000-card domestic ASIC cluster, no NVIDIA) — 1.6T MoE with dynamic per-token compute (33B–56B active) via zero-computation experts.
PriceWeights free (MIT); API list ~$0.75 / 1M input, ~$2.95 / 1M output; launch promo ~$0.30 / ~$1.20; cached input free
DeploymentLocal / self-host (HF); LongCat API Platform; OpenRouter; vLLM, SGLang (GPU + NPU), llama.cpp / Ollama / LM Studio (quantized)
ModalitiesText-only
ParametersMoE 1.6T total / ~48B active per token (dynamic 33B–56B; zero-computation experts + PID-balanced activation); ScMoE shortcut-connected design; 135B-param N-gram embedding; LongCat Sparse Attention (streaming-aware, cross-layer, hierarchical indexing); native 1M context, 128K max output; pretrained on 35T+ tokens; thinking via enable_thinking; released 30 Jun 2026 (codename "Owl Alpha" before launch)
ProsMIT open weights; near-frontier agentic coding; very cheap API (cached reads free); OpenAI- and Anthropic-compatible endpoints; native 1M context; vLLM / SGLang support incl. NPU path
ConsEnormous to self-host (1.6T weights, multi-node); text-only; vendor-reported benches; launch price is a promo with no end date; behind Kimi K3 / GPT-5.6 Sol on Terminal-Bench 2.1
Modelmeituan-longcat/LongCat-2.0 (API: LongCat-2.0; OpenRouter: meituan/longcat-2.0)
Scalelarge / frontier
Benchmarking- SWE-Bench Pro59.5%
- Terminal-Bench 2.170.8%
- SWE-bench Multilingual77.3%
- BrowseComp79.9%
- GPQA-Diamond88.9%
Source: Meituan-reported (huggingface.co/meituan-longcat/LongCat-2.0 card, MarkTechPost launch coverage). SWE-bench Pro 59.5 vs GPT-5.5 58.6. Not a third-party rerun.
- Muse Spark 1.1 (Meta)Meta (Muse)Reasoning / thinker
Released9 Jul 2026
Best forAgentic tool use (MCP), computer use (desktop / browser agents), professional tool workflows, coding agents, multimodal reasoning over images, video, charts and PDFs, 1M-context work. Unique: Meta's first closed, hosted frontier model (Meta Superintelligence Labs, not Llama) — tops Claude Opus 4.8 and GPT-5.5 on MCP Atlas and JobBench; computer-use agent picks per step between writing a script and clicking the UI.
Price~$1.25 / 1M input, ~$4.25 / 1M output; $20 starter credits; free on meta.ai (consumer)
DeploymentAPI only (Meta Model API public preview, OpenRouter); Meta AI app, meta.ai
ModalitiesText + image + video + audio + PDF
ParametersParameter count / architecture not published; 1M-token context; multimodal reasoning model with Thinking mode; built-in tool use + computer use; released 9 Jul 2026 (first Muse Spark Apr 2026; Muse Spark 1.3 out 2 Sep 2026)
ProsBest-in-class tool use and strong computer use; 1M context; native multimodal input; cheap for frontier tier; free consumer access on meta.ai
ConsClosed weights (no self-host, unlike Llama); no published size / architecture; behind Opus 4.8 and GPT-5.5 on long-horizon coding and Terminal-Bench 2.1; API preview US-only at launch; superseded by Muse Spark 1.3
ModelMeta Model API (exact id: confirm in Meta console); OpenRouter: meta/muse-spark-1.1
Scalelarge / frontier
BenchmarkingSource: Meta-reported (Muse Spark 1.1 evaluation report). Comparisons vs Claude Opus 4.8 / GPT-5.5: MCP Atlas 88.1 (82.2 / 75.3); JobBench 54.7 (48.4 / 38.3); OSWorld 80.8 (83.4 / 78.7); DeepSWE 1.1 53.3 (59.0 / 67.0); Terminal-Bench 2.1 80.0 (82.7 / 83.4); SWE-Bench Pro 61.5 (69.2 / 58.6). Not a third-party rerun.
- Laguna S 2.1 (Poolside)Poolside (Laguna)Reasoning / thinker
Released21 Jul 2026
Best forAgentic coding and long-horizon software engineering (SWE, terminal agents, multilingual repos, codebase Q&A), tool-using coding agents on local / single-node hardware, 1M-context repo work. Unique: small-active MoE (8B active of 118B) that matches or beats open models several times its size on agentic coding.
PriceWeights free (OpenMDW-1.1); infra cost only; also hosted on third-party platforms
DeploymentLocal / self-host (HF); vLLM, SGLang, TRT-LLM, llama.cpp, Ollama, LM Studio (FP8, NVFP4, INT4, GGUF)
ModalitiesText-only
ParametersMoE 118B total / 8B active; 256 routed experts + 1 shared, softplus-gated token-choice router; 48 layers (12 global + 36 sliding-window, window 512); GQA (8 KV heads); 1,048,576-token context; interleaved thinking between tool calls (enable_thinking); DFlash draft model for speculative decoding; released ~21 Jul 2026 (family: Laguna XS 2.1 33B-A3B, Laguna M.1 225B-A23B)
ProsPermissive open weights (OpenMDW-1.1, commercial use OK); strong coding for 8B active; 1M context; easy to serve with many quantized formats; cheap per-token compute
ConsText-only; below Kimi K3 / Claude Fable 5 on hard agent benches; still ~118B weights to load; vendor-reported benches; newer license — check terms
Modelpoolside/Laguna-S-2.1
Scalemedium
BenchmarkingSource: Poolside-reported (huggingface.co/poolside/Laguna-S-2.1 card). Also: SWE Atlas (Codebase QnA) 46.2%; Toolathlon Verified 49.7%. Not a third-party rerun.
- Metis (GreenPT)GreenPTClassification / decision
Released8 Oct 2026
Best forFast, typed decisions inside programs for EU-based teams — ticket triage, request sorting, risk flags, priority setting, yes/no gates, agent step choices. Unique: European decision engine (like JEV) — one pass returns a structured answer (yes / no, multiple choice or scale rating) with a confidence score for every option, served only from renewable-powered EU data centers, with an energy-usage report per request.
PriceNot disclosed at launch
DeploymentAPI only (EU data centers)
ModalitiesText-only (structured question + context in, typed answer + confidences out)
ParametersParameter count / architecture not published; decision model, not a text generator — processes input in a single pass instead of token-by-token generation; output limited to structured types (yes / no, multiple choice, scale) with per-option confidence; typical latency < 100 ms (72 ms on a support-ticket example); up to 32 questions per request
ProsVery low latency (< 100 ms, vendor claims up to 5× faster than chat-based systems); always returns a valid typed answer with calibrated confidences; batches up to 32 questions; EU data residency, customer data not used for training; per-request energy reporting
ConsClosed / API-only; no published pricing, size or standard benchmarks; brand-new (launched 8 Oct 2026) with no independent evaluations yet; no free-text output (can't explain, write or chat); smaller ecosystem than JEV (no Vercel AI Gateway / benchmark coverage yet)
ModelMetis (GreenPT API; exact model id: confirm in GreenPT docs)
ScaleN/A
BenchmarkingNo public benchmark scores at launch; vendor claims < 100 ms latency and up to 5× faster than chat-based systems; positioned against JEV (TypeSafe AI)
- Nex-N2-Pro (Nex AGI)Nex AGIReasoning / thinker
Released2 Jun 2026 (open weights); listed on OpenRouter ~8 Jun 2026
Best forAgentic coding and tool-using agents (SWE, terminal, function calling, structured output), reasoning-heavy work, image + text understanding; self-hosted open alternative to closed coding models. Unique: "Agentic Thinking" post-training — adaptive + coherent thinking that unifies reasoning, tool use and environment execution in one loop — on top of Qwen3.5-397B, beating GPT-5.5 on SWE-Bench Pro and Claude Opus 4.7 on Terminal-Bench 2.1.
PriceWeights free (Apache 2.0); hosted via OpenRouter (see current pricing)
DeploymentLocal / self-host (HF, ModelScope); Nex SGLang fork (recommended), vLLM, SGLang, Transformers, Docker; llama.cpp / Ollama / LM Studio (quantized); OpenRouter
ModalitiesText + image in, text out
ParametersMoE 397B total / 17B active per token; post-trained on Qwen3.5-397B-A17B (not trained from scratch); 262,144-token context; BF16 weights; function calling (`qwen3_coder` parser) + reasoning parser (`qwen3`); recommended `temperature=0.7`, `top_p=0.95`, `top_k=40`
ProsApache 2.0 open weights (HF + ModelScope); strong coding for open models (SWE-Bench Pro 58.8 vs GPT-5.5 58.6; Terminal-Bench 2.1 75.3 vs Opus 4.7 69.7); text + image input; many quantized builds (llama.cpp, Ollama, LM Studio); OpenRouter access
ConsPost-train of Qwen3.5, not a new base; behind Opus 4.7 and GPT-5.5 on SWE-Bench Verified and GPQA; weak on DeepSWE (33.6%); best served with Nex's custom SGLang fork; vendor-reported benches
Modelnex-agi/Nex-N2-Pro (OpenRouter: nex-agi/nex-n2-pro)
Scalelarge / frontier
BenchmarkingSWE-Bench Verified 80.8 (82.9 / 87.6); SWE-Bench Pro 58.8 (58.6 / 64.3); Terminal-Bench 2.1 75.3 (83.4 / 69.7); GPQA Diamond 90.7 (93.6 / 94.2); GDPval 1585 (1769 / 1753); DeepSWE 33.6% (third-party listing)
- Rio-3.5-Open-397B (IplanRIO)IplanRIO (Rio de Janeiro)Reasoning / thinker (merged model)
Released13 Jun 2026
Best forMainly a case study in open-model provenance and attribution, not a recommended pick — for real use, take the source models directly (Nex-N2-Pro for agentic coding, Qwen3.5-397B for a general base). Unique: marketed as a "Brazilian" frontier model by Rio de Janeiro's municipal IT company, but analysis showed the uploaded weights are a layer-wise blend of ~60% Nex-N2-Pro + ~40% Qwen3.5-397B-A17B across all 60 layers.
PriceWeights free (MIT per card); infra cost only
DeploymentLocal / self-host (HF); Qwen3.5-compatible engines (vLLM, SGLang); community quantizations (e.g. NVFP4)
ModalitiesText-only (per listings; Nex-N2-Pro base also takes images — confirm on current card)
ParametersMoE 397B total / ~17B active (same architecture as Qwen3.5-397B-A17B — not a new architecture); weights = layer-wise merge of Nex-N2-Pro and Qwen3.5-397B, per Nex AGI and Hugging Face researcher Wang Tiezhen; IplanRIO's updated card claims merge + on-policy distillation from a stronger model and says the first upload was the wrong checkpoint (a merged baseline used for comparison); community NVFP4 quantization exists
ProsOpen weights (MIT per card); Terminal-Bench 2.1 70.8 (vendor-reported; above DeepSeek-V4-Pro 67.9 and base Qwen3.5 52.5); runs on the same Qwen3.5 serving stack (vLLM, SGLang, quantized builds)
ConsProvenance dispute — weights match a Nex + Qwen merge, not independent training; IplanRIO apologized (15 Jun 2026) for an "operational error" and promised a re-upload of the distilled version; benchmarks need independent audit; licensing / credit to Nex AGI and Alibaba contested; Nex-N2-Pro itself scores higher on Terminal-Bench 2.1 (75.3)
Modelprefeitura-rio/Rio-3.5-Open-397B (Hugging Face; confirm current repo and checkpoint version)
Scalelarge / frontier
BenchmarkingTerminal-Bench 2.1 70.8 (DeepSeek-V4-Pro 67.9; Qwen3.5-397B 52.5)
- MiniCPM5-1B (OpenBMB)OpenBMBReasoning / thinker
Released19 May 2026
Best forLocal / edge use on phones and laptops — light coding, tool calling, short reasoning, offline assistants, privacy-sensitive apps. Unique: 1B-class open-source SOTA per OpenBMB, with the biggest lead in tool use, code generation and hard reasoning; Artificial Analysis Intelligence Index 17.9 (non-reasoning), top open-weight sub-2B model at the time.
PriceFree (open weights, Apache 2.0); infra / device cost only
DeploymentLocal / on-device (phone, laptop, CPU); llama.cpp, Ollama, vLLM, MLX; Hugging Face
ModalitiesText-only
Parameters~1B params; Llama-style dense architecture; 128K context; optional think / no-think mode; Apache 2.0
ProsApache 2.0 open weights; runs fully on-device (llama.cpp, Ollama, vLLM, MLX); 128K context at 1B size; thinking mode for harder prompts; beats Qwen3-0.6B, Qwen3.5-0.8B and LFM2.5-1.2B on OpenBMB's benchmark set
ConsLow ceiling vs larger models (fine for light tasks, not complex agents); benchmark average is vendor-selected; text-only; MiniCPM5-2B is noticeably stronger if 2B fits your device
Modelopenbmb/MiniCPM5-1B
Scaletiny
Benchmarkingselected-benchmark average 42.57 (Thinking) vs Qwen3-0.6B Thinking 26.77, Qwen3.5-0.8B Thinking 25.14, LFM2.5-1.2B 35.61; Artificial Analysis Intelligence Index 17.9 (non-reasoning, third-party, 26 May 2026)
- Devstral 2 (Mistral AI)Mistral AIGenerative (code-focused)
Released9 Dec 2025
Best forAgentic coding — SWE tasks, multi-file edits, repo-level changes in a terminal agent (Mistral Vibe CLI, Zed); self-hosted coding agent, incl. laptop-class via Devstral Small 2. Unique: coding-agent model pair shipped together with Mistral's own open-source terminal agent (Vibe CLI); the small 24B sibling runs on a single GPU or laptop offline.
PriceWeights free; API ~$0.40 / 1M input, ~$2.00 / 1M output (Devstral 2); ~$0.10 / ~$0.30 (Devstral Small 2); free API period at launch
DeploymentMistral API; local / self-host (HF, Ollama, llama.cpp); Mistral Vibe CLI; Zed extension
ModalitiesText-only
ParametersDevstral 2 = 123B dense transformer, 256K context; Devstral Small 2 = 24B dense, 256K context; tuned for agentic software engineering (tool use, multi-file orchestration); Devstral 2 needs ≥4 H100-class GPUs, Small 2 runs on one GPU (RTX, DGX Spark) or CPU
ProsOpen weights (Devstral 2: modified MIT; Small 2: Apache 2.0); strong SWE-bench Verified for size (72.2% / 68.0%); cheap API ($0.40 / $2.00; Small 2 $0.10 / $0.30); Small 2 fully local and private; human eval preferred over DeepSeek V3.2 (42.8% win vs 28.6% loss)
ConsDeprecated on Mistral's API by Sep 2026 — new integrations pointed to Mistral Medium 3.5; behind Claude Sonnet 4.5 in Mistral's own human eval; older than the 2026 coding leaders; modified MIT license has extra terms for Devstral 2 (check before commercial use); text-only
ModelAPI: devstral-medium-latest (Devstral 2), Devstral Small 2 via devstral-small id (confirm in Mistral docs); HF: mistralai/Devstral-2-123B-Instruct-2512, mistralai/Devstral-Small-2-24B-Instruct-2512 (confirm)
Scalelarge
BenchmarkingSWE-bench Verified 72.2% (Devstral 2), 68.0% (Devstral Small 2); human eval vs DeepSeek V3.2: 42.8% win / 28.6% loss
- Hy3 (Tencent Hunyuan)Tencent HunyuanReasoning / thinker
Released6 Jul 2026 (official); preview 23 Apr 2026
Best forReasoning, coding and agent work at mid-size cost — SWE tasks, office documents (PPT / Word / Excel / PDF / HTML), knowledge-base Q&A, multi-agent workflows; self-hosting on a single 8-GPU node. Unique: Tencent's rebuilt Hunyuan flagship — hybrid fast / slow-thinking MoE with only 21B active that Tencent says matches flagships 2–5× its size; runs inside Tencent products (Yuanbao, CodeBuddy, WorkBuddy, Weixin) at scale.
PriceWeights free (Apache 2.0); API ~¥1.00 / 1M input, ~¥0.25 / 1M cached input, ~¥4.00 / 1M output (≈ $0.14 / $0.56)
DeploymentLocal / self-host (HF, ModelScope, GitCode); vLLM, SGLang (8 GPUs, H20-3e class or better); Tencent Cloud TokenHub; OpenRouter
ModalitiesText-only
ParametersMoE 295B total / 21B active; 192 experts, top-8 per token; extra multi-token-prediction (MTP) layer of 3.8B params; hybrid fast / slow thinking; 256K context; text-only; BF16 and FP8 checkpoints
ProsApache 2.0 open weights; strong for 21B active (SWE-bench Verified 78%, GPQA Diamond 90.4%, HLE 53.2%); cheap API (¥1 / ¥4 per 1M tokens, cached ¥0.25); single-node self-host (8× H20-class GPUs); wide tool support (OpenRouter, Cline, Kilo, OpenCode, Cherry Studio)
ConsText-only; weak on long-horizon terminal tasks (Long-Horizon Terminal Bench 28.8%); vendor-reported benches (preview and official numbers differ, e.g. SWE-bench Verified 74.4% preview vs 78% official); API priced in CNY via Tencent Cloud
Modeltencent/Hy3 (also tencent/Hy3-FP8); API (Tencent Cloud TokenHub): hy3
Scalelarge
BenchmarkingSWE-bench Verified 78%; SWE-bench Pro 57.9%; GPQA Diamond 90.4%; HLE 53.2%; SkillsBench v1.1 55.3%; Long-Horizon Terminal Bench 28.8%
- Grok 4.7 (xAI)xAIReasoning / thinker
Released21 Sep 2026
Best forLong-horizon coding and debugging, agentic knowledge work, legal and electrical-engineering agent tasks, documents and presentations; cheap frontier-class API for coding agents (Cursor, Grok Build). Unique: top score on legal agent work (Harvey Legal Agent 19.6% vs Fable 5.1 6.7% and GPT-5.6 Sol 2.5%) and electrical engineering (EEBench 64.0%) at ~1/8 of Claude Fable 5.1's output price.
Price~$2 / 1M input, ~$0.50 / 1M cached input, ~$6 / 1M output (≤200K prompt); ~$4 / ~$12 above 200K; fast variant ~$4 / ~$12
DeploymentAPI only (xAI API, Microsoft Foundry, Amazon Bedrock, Oracle Cloud); Cursor; Grok Build
ModalitiesText-only (coding and knowledge work)
ParametersParameter count not published (new, larger base model than Grok 4.6); longer RL run on hard, many-hour tasks; 500K context (per Artificial Analysis — not stated in xAI's launch post); reasoning effort levels (benchmarks at xhigh); fast variant at 2× speed; text-focused; closed weights
ProsCheap for frontier tier ($2 / $6, cached $0.50, unchanged from 4.6); strong long-horizon coding (DeepSWE v1.1 71.0%, above Fable 5.1 70.0%); big jump vs 4.6 (Terminal-Bench 4.0 20.3% → 38.0%, CursorBench 40.4% → 46.3%); wide availability (xAI API, Cursor, Grok Build free tier, Microsoft Foundry, Amazon Bedrock, Oracle Cloud)
ConsClosed weights; behind Fable 5.1 on CursorBench (46.3% vs 51.8%) and Terminal-Bench 4.0 (Artificial Analysis measures 26% vs xAI's 38%); context size not confirmed by xAI; price doubles above 200K prompt tokens; Grok consumer apps still serve 4.6
Modelgrok-4.7 (Amazon Bedrock: us.xai.grok-4.7; Oracle Cloud: xai.grok-4.7)
Scalelarge / frontier
BenchmarkingDeepSWE v1.1 71.0% (70.0 / 72.7); CursorBench 4.0 46.3% (51.8 / 41.7); Terminal-Bench 4.0 38.0% (xAI harness; Artificial Analysis 26%); Harvey Legal Agent 19.6% (6.7 / 2.5); EEBench 64.0% (56.4 / 39.4); HealthBench Professional 56.7% (62.1 / 60.5); Coding Agent Index 56 with Grok Build; AA Intelligence Index 46 (third-party)
- Nemotron 3 Super (NVIDIA)NVIDIAReasoning / thinker
Released11 Mar 2026
Best forFast, cheap open model for multi-agent and coding workflows, very long-context work (1M tokens), enterprise self-hosting on NVIDIA GPUs; teams that need fully open training data and recipes. Unique: hybrid Mamba-2 + Transformer "LatentMoE" — routes tokens in a compressed 1,024-dim latent space so it can use 512 experts (22 active) with only 12B active params; fastest model in its class (478 tok/s per Artificial Analysis) and 91.75% on RULER at 1M tokens.
PriceWeights free (NVIDIA Open Model License); API ~$0.10 / 1M input, ~$0.50 / 1M output (DeepInfra); free tier on OpenRouter
DeploymentLocal / self-host (HF); NVIDIA NIM; vLLM / SGLang / TensorRT-LLM; DeepInfra, OpenRouter and other API providers
ModalitiesText-only
ParametersMoE 120B total / 12B active; 88-layer hybrid (Mamba-2 layers + periodic Transformer attention layers as "global anchors"); LatentMoE: hidden 4,096 projected to 1,024 latent before routing, 512 experts, 22 active per token; built-in multi-token prediction (avg 3.45 tokens accepted per step, up to ~3× speedup); 1M context (APIs often cap at 262K); text-only; pretrained on 25T tokens in NVFP4 from the first step; post-training ~7M SFT samples + RL on 21 environments (1.2M rollouts)
ProsOpen weights + released datasets and full pretraining / SFT / RL recipes; very fast (2.2× GPT-OSS-120B throughput); 1M context with strong recall (RULER 1M 91.75% vs GPT-OSS-120B 22.3%); very cheap API (~$0.10 / $0.50, free tier on OpenRouter); BF16, FP8 and NVFP4 checkpoints; NVIDIA NIM support; PinchBench 85.6% (best open model at launch)
ConsNVIDIA Open Model License — its own terms for distribution, not MIT / Apache (check before redistributing); weaker coding than Qwen3.5-122B (SWE-bench Verified 60.47% vs 66.4%); behind GPT-OSS-120B on Arena-Hard V2; text-only; full 1M context needs self-hosting (8× H100-80GB); superseded at the top of the family by Nemotron 3 Ultra
Modelnvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 (also -FP8, NVFP4, base); OpenRouter: nvidia/nemotron-3-super-120b-a12b
Scalelarge
BenchmarkingMMLU-Pro 83.73% (86.70 / 81.00); SWE-bench Verified 60.47% (66.40 / 41.90); RULER @ 1M 91.75% (— / 22.30); Arena-Hard V2 73.88% (— / 90.26); GPQA (with tools) 82.70% (— / 80.09); PinchBench 85.6%; throughput 478 tok/s (Artificial Analysis)
- Ling-3.1-flash (Ant Group / inclusionAI)Ant Group (inclusionAI)Reasoning / thinker
Released30 Sep 2026 (announced + hosted preview; OpenRouter listing 1 Oct 2026)
Best forReal-world agent work in finance, healthcare and cybersecurity, codebase Q&A, long-horizon coding, front-end / SVG generation; free hosted trial for evaluating a large Chinese MoE. Unique: Ant Group's biggest "flash" model — ~560B MoE with ~25B active that tops Ant's own chart on FrontierSWE (75.16), HealthBench Professional (65.35) and CyberGym (87.9); hybrid reasoning across a 1M-token context.
PriceFree during launch trial (OpenRouter / Novita); paid pricing and open-weight license not announced
DeploymentAPI only for now (OpenRouter via Novita, OpenCode); self-hosting after the promised open-weight release
ModalitiesText-only
ParametersMoE ~560B total / ~25B active per token; hybrid reasoning; up to 1M context (262,144 during hosted trial); 32,768-token output limit on OpenRouter / Novita; text-only; predecessor Ling-3.0-flash was 124B / 5.1B active (Aug 2026, MIT)
ProsStrong vendor scores on agentic work (GDPval-AA v2.1 Elo 1673; FrontierSWE 75.16; DRACO 85.49); leads on HealthBench Professional and CyberGym in Ant's chart; free during the hosted trial; 1M context; hands-on tests good at SQL optimization, front-end / SVG and clinical reasoning
ConsNot open yet — no Hugging Face repo, weights or license as of mid-Oct 2026 (only "plan to open-source soon"); all benchmarks Ant-reported, none reproduced; Claude Opus 5 leads on several tests incl. terminal tasks (Terminal-Bench 4.0 40.4%); weak on complex multi-tool pipelines and rare languages in hands-on tests; free trial was time-limited (~2 weeks)
ModelOpenRouter: inclusionai/ling-3.1-flash (served by Novita); HF repo: not published yet
Scalelarge / frontier
BenchmarkingGDPval-AA v2.1 Elo 1673; FrontierSWE 75.16; HealthBench Professional 65.35; CyberGym 87.9; DRACO 85.49; Terminal-Bench 4.0 40.4