{"ok":true,"source":"tensorfeed.ai","lastUpdated":"2026-09-14","count":24,"models":[{"id":"codestral-25.08","name":"Codestral 25.08","publisher":"Mistral","domain":"code","params":"undisclosed","pricing":"$0.30 input / $0.90 output per 1M tokens via Mistral API","openWeights":false,"license":"Mistral commercial (Premier model)","released":"2025-07","benchmark":null,"capabilities":["code completion","fill-in-the-middle","chat + function calling","128k context"],"url":"https://docs.mistral.ai/models/codestral-25-08","notes":"Mistral's low-latency completion model and the current Codestral; Codestral 25.01 was retired from the API on 2025-11-30. Mistral reports +30% accepted completions and 50% fewer runaway generations over prior versions but publishes no absolute benchmark scores. Deployable in cloud, VPC, or on-prem under enterprise terms. For agentic coding, Mistral now routes former Devstral users to Mistral Medium 3.5."},{"id":"mai-code-1.1-flash","name":"MAI-Code-1.1-Flash","publisher":"Microsoft AI","domain":"code","params":"138B MoE / 5B active","pricing":"$0.20 input / $1.20 output per 1M tokens (GitHub Copilot list rate)","openWeights":false,"license":"Proprietary (GitHub Copilot terms)","released":"2026-08","benchmark":"SWE-Bench Verified 72.6%, Terminal-Bench 2.1 62.9% (Copilot harness, vendor-run)","capabilities":["agentic coding","image input","256k context","GitHub Copilot integration"],"url":"https://github.com/microsoft/MAI-Code","notes":"Microsoft AI's in-house small-tier coding model, trained against the GitHub Copilot production harness and distributed only inside Copilot clients. Succeeds MAI-Code-1-Flash (June 2026) at a quarter of its list price and adds vision input. All published scores come from Microsoft's own Copilot harness, so treat cross-lab comparisons with care."},{"id":"qwen-coder-2.5-32b","name":"Qwen 2.5 Coder 32B","publisher":"Alibaba","domain":"code","params":"32B","pricing":"Open weights, free to self-host","openWeights":true,"license":"Apache-2.0","released":"2024-11","benchmark":"HumanEval 92.7%, MBPP 90.2%","capabilities":["code completion","code repair","40+ languages","128k context"],"url":"https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct","notes":"Apache-2.0 dense code model that punches above its parameter count. Still a solid self-hosted default, though the Qwen line has moved on to Qwen3-Coder (480B-A35B and 30B-A3B) and Qwen3-Coder-Next (80B MoE / 3B active, Apache-2.0, Feb 2026) for agentic coding."},{"id":"starcoder-2-15b","name":"StarCoder 2 15B","publisher":"BigCode","domain":"code","params":"15B","pricing":"Open weights, free to self-host","openWeights":true,"license":"BigCode OpenRAIL-M","released":"2024-02","benchmark":"HumanEval 46.3% pass@1","capabilities":["code completion","fill-in-the-middle","600+ languages","16k context"],"url":"https://huggingface.co/bigcode/starcoder2-15b","notes":"Reproducible: trained on The Stack v2 (open dataset). Best fit for academic / regulated code-completion deployments."},{"id":"med-gemini","name":"Med-Gemini","publisher":"Google","domain":"medical","params":"undisclosed","pricing":"Not publicly available (research model)","openWeights":false,"license":"Proprietary","released":"2024-04","benchmark":"MedQA-USMLE 91.1%","capabilities":["clinical Q&A","medical imaging","EHR summarization","long-context records"],"url":"https://research.google/blog/advancing-medical-ai-with-med-gemini/","notes":"Google's research family of Gemini models fine-tuned for medicine (radiology, pathology, dermatology, genomics). Set a state-of-the-art 91.1% on MedQA at publication. Google says considerable further research is needed before real-world clinical use, so it is not a self-serve product; Google's downloadable medical line is MedGemma."},{"id":"apertus-70b-meditronfo","name":"Apertus-70B-MeditronFO","publisher":"EPFL (LiGHT lab)","domain":"medical","params":"70B","pricing":"Open weights, free to self-host","openWeights":true,"license":"Apache-2.0","released":"2026-05","benchmark":"MedQA 68.6%, HealthBench Hard 40.1%","capabilities":["clinical Q&A","guideline-grounded answers","differential diagnosis"],"url":"https://huggingface.co/EPFLiGHT/Apertus-70B-MeditronFO","notes":"Flagship of EPFL's Fully Open Meditron family: open weights, open training data, and an open recipe, fine-tuned from Swiss AI's Apertus-70B-Instruct on a clinician-vetted corpus built partly from 46,469 clinical practice guidelines. Siblings ship on Apertus 8B, Gemma 3 27B, EuroLLM, and OLMo 2 bases. Succeeds the Llama 2 based Meditron 70B as EPFL's reference open medical model."},{"id":"biomistral","name":"BioMistral 7B","publisher":"BioMistral collaboration","domain":"medical","params":"7B","pricing":"Open weights, free to self-host","openWeights":true,"license":"Apache-2.0","released":"2024-02","benchmark":"MedQA-USMLE 50.6% (4-option)","capabilities":["biomedical Q&A","drug interactions","literature retrieval"],"url":"https://huggingface.co/BioMistral/BioMistral-7B","notes":"Mistral 7B continued-pretrained on PubMed. Apache-licensed; small enough to self-host on a single consumer GPU. Good base for medical RAG agents."},{"id":"thomson-1.0-small","name":"Thomson-1.0-Small","publisher":"Thomson Reuters","domain":"legal","params":"35B MoE / 3B active","pricing":"Open weights, free for non-commercial use","openWeights":true,"license":"PolyForm Strict 1.0.0 (non-commercial)","released":"2026-08","benchmark":"Harvey Legal Agent Benchmark 73.4%, LegalBench 79.9% (vendor harness)","capabilities":["legal research agents","contract understanding","tax Q&A","262k context"],"url":"https://huggingface.co/thomsonreuters/Thomson-1.0-Small","notes":"Open-weight member of Thomson Reuters' Thomson-1.0 family, continually trained from Qwen3.6-35B-A3B on proprietary case law, statutes, contracts, filings, and news. The larger proprietary Thomson model powers Tabular Analysis in CoCounsel Legal. Gains show up on agentic and document tasks; on plain LegalBench it lands slightly below its Qwen base."},{"id":"saul-lm-141b","name":"SaulLM 141B","publisher":"Equall","domain":"legal","params":"141B (Mixtral-based)","pricing":"Open weights, free to self-host","openWeights":true,"license":"MIT","released":"2024-07","benchmark":"Beat prior open models on LegalBench-Instruct at release (paper)","capabilities":["legal Q&A","contract review","case law summarization","multilingual legal"],"url":"https://huggingface.co/Equall/SaulLM-141B-Instruct","notes":"Largest open legal LLM at its release. Continued pretraining of Mixtral on a 540B+ token legal corpus, released in base, instruct, and aligned versions. Trained on US and European legal texts; a research artifact, not legal advice."},{"id":"saul-lm-7b","name":"SaulLM 7B","publisher":"Equall","domain":"legal","params":"7B","pricing":"Open weights, free to self-host","openWeights":true,"license":"MIT","released":"2024-03","benchmark":null,"capabilities":["legal Q&A","contract clauses","case briefing"],"url":"https://huggingface.co/Equall/Saul-7B-Base","notes":"Smaller SaulLM, continued pretraining of Mistral 7B on 30B+ tokens of English legal text. Runs on consumer GPU at Q4. Solid base for legal RAG agents."},{"id":"fingpt-v3","name":"FinGPT v3","publisher":"AI4Finance Foundation","domain":"finance","params":"6B-13B (LoRA on ChatGLM2 / Llama 2)","pricing":"Open weights, free to self-host","openWeights":true,"license":"MIT","released":"2023-10","benchmark":"FPB weighted F1 0.882 (v3.3)","capabilities":["financial sentiment","news + tweet classification","single-GPU LoRA fine-tuning"],"url":"https://github.com/AI4Finance-Foundation/FinGPT","notes":"Open finance models maintained by AI4Finance Foundation. v3 is a set of LoRA adapters: v3.1 on ChatGLM2-6B, v3.2 on Llama 2 7B, v3.3 on Llama 2 13B, trainable on a single RTX 3090. Best fit for sentiment + classification rather than reasoning."},{"id":"bloomberggpt","name":"BloombergGPT","publisher":"Bloomberg","domain":"finance","params":"50B","pricing":"Internal Bloomberg use only (not public)","openWeights":false,"license":"Proprietary","released":"2023-03","benchmark":"FPB weighted F1 51.07 (5-shot)","capabilities":["financial NER","sentiment","headline classification","financial QA (ConvFinQA)"],"url":"https://www.bloomberg.com/company/press/bloomberggpt-50-billion-parameter-llm-tuned-finance/","notes":"Bloomberg's in-house finance LLM. Trained on a private 363B-token finance corpus plus public data. Reference for \"what does a finance-specialized base model look like\" but not externally usable."},{"id":"suno-v6","name":"Suno v6","publisher":"Suno","domain":"music","params":"undisclosed","pricing":"Free tier (v6-mini); Pro $8/mo, Premier $24/mo billed yearly","openWeights":false,"license":"Proprietary","released":"2026-09","benchmark":null,"capabilities":["lyrics + music","genre control","v6-wild exploratory variant","free v6-mini tier"],"url":"https://suno.com/blog/introducing-v6","notes":"Suno's first model generation developed with industry partners including Warner Music Group, BMG, and Believe. Ships as v6 (flagship, paid), v6-wild (less predictable, paid), and v6-mini (free). Suno is retiring its earlier models as v6 rolls out. The default for music-gen agents that need polished output."},{"id":"udio","name":"Udio","publisher":"Uncharted Labs","domain":"music","params":"undisclosed","pricing":"Subscription plans (see udio.com/pricing)","openWeights":false,"license":"Proprietary","released":"2024-04","benchmark":null,"capabilities":["music gen","extend","remix","lyrics or instrumental"],"url":"https://www.udio.com","notes":"Suno competitor that settled with Universal Music Group (Oct 2025) and Warner Music Group (Nov 2025). Since the UMG deal, creations stay inside a walled garden and cannot be downloaded. A licensed subscription platform trained on authorized music is in development, so it is a poor fit for agents that need exportable audio today."},{"id":"minimax-music-3","name":"MiniMax Music 3","publisher":"MiniMax","domain":"music","params":"8B global LM + 0.6B local LM + 2.4B flow-matching decoder","pricing":"Open weights, free to self-host","openWeights":true,"license":"MiniMax-Music3 Community License","released":"2026-08","benchmark":null,"capabilities":["lyrics + music","full songs up to 5 min","structured style captions","32 kHz stereo output"],"url":"https://huggingface.co/MiniMaxAI/MiniMax-Music3","notes":"Open full-song model with vocals: an 8B global LLM (initialized from Qwen3-8B) plans structure, a 0.6B local LLM fills acoustic detail, and a flow-matching decoder renders audio. Fits under 24GB VRAM, or 8GB with offloading. Commercial use requires on-screen attribution, and products above $20M yearly revenue need written authorization."},{"id":"musicgen-large","name":"MusicGen Large","publisher":"Meta","domain":"music","params":"3.3B","pricing":"Open weights, free to self-host","openWeights":true,"license":"CC-BY-NC-4.0","released":"2023-06","benchmark":null,"capabilities":["instrumental music gen","melody conditioning","audio continuation"],"url":"https://huggingface.co/facebook/musicgen-large","notes":"Open instrumental music model trained on licensed Meta, Shutterstock, and Pond5 music. Non-commercial weights license. Now an older research baseline; newer open options include Stable Audio 3.0 and MiniMax Music 3."},{"id":"stable-audio-3","name":"Stable Audio 3.0","publisher":"Stability AI","domain":"music","params":"1.4B (Medium) / 2.7B (Large) diffusion transformer","pricing":"Small + Medium open weights; Large via API at 26 credits ($0.26) per generation","openWeights":true,"license":"Stability AI Community License (Enterprise License above $1M revenue)","released":"2026-05","benchmark":null,"capabilities":["music + SFX gen","up to 6m20s clips","audio-to-audio","inpainting / continuation"],"url":"https://stability.ai/news-updates/meet-stable-audio-3-the-model-family-built-for-artistic-experimentation-with-open-weight-models","notes":"Stability's audio family trained on licensed AudioSparx audio plus filtered Creative Commons Freesound recordings. Small variants run on CPU and phones, Medium needs a CUDA GPU, Large is API and enterprise only. Designed for sound design workflows; replaces Stable Audio 2.0, which is no longer on Stability's API price list."},{"id":"trellis","name":"TRELLIS","publisher":"Microsoft Research","domain":"3d","params":"up to 2B","pricing":"Open weights, free to self-host","openWeights":true,"license":"MIT","released":"2024-12","benchmark":null,"capabilities":["image to 3D","text to 3D","mesh + gaussian splat output","local 3D editing"],"url":"https://github.com/microsoft/TRELLIS","notes":"Microsoft Research 3D generative model built on a unified structured latent. Outputs textured mesh, Gaussian splats, or radiance fields. For image-to-3D, Microsoft's TRELLIS.2 (4B, MIT, Dec 2025) is the successor and adds full PBR materials."},{"id":"hunyuan3d-2","name":"Hunyuan3D-2","publisher":"Tencent","domain":"3d","params":"1.1B shape + 1.3B texture","pricing":"Open weights, free to self-host","openWeights":true,"license":"Tencent Hunyuan Community License","released":"2025-01","benchmark":null,"capabilities":["image to 3D mesh","texture generation","text to 3D"],"url":"https://huggingface.co/tencent/Hunyuan3D-2","notes":"Tencent's open 3D model. Two-stage: shape DiT + texture paint model. Strongest open 3D output quality in early 2025. Hunyuan3D-2.1 (June 2025) followed with PBR texture synthesis and released training code."},{"id":"colpali","name":"ColPali","publisher":"ILLUIN Technology","domain":"retrieval","params":"3B (PaliGemma-based)","pricing":"Open weights, free to self-host","openWeights":true,"license":"MIT","released":"2024-07","benchmark":"ViDoRe avg nDCG@5 81.3 (paper)","capabilities":["vision-document retrieval","late interaction","PDF / chart understanding"],"url":"https://huggingface.co/vidore/colpali","notes":"Vision-document retrieval model. Embeds whole-page images directly, skipping OCR. Strong on PDFs with tables, charts, scientific figures. Newer checkpoints (colpali-v1.3, ColQwen2.5) live under the same vidore org."},{"id":"evie-8b","name":"EVIE-8B","publisher":"Tencent","domain":"retrieval","params":"8.4B (Qwen3.5-based)","pricing":"Open weights, free to self-host","openWeights":true,"license":"Apache-2.0","released":"2026-09","benchmark":"ViDoRe V3 nDCG@10 66.75, ViDoRe V1 nDCG@5 92.18 (vendor-reported)","capabilities":["vision-document retrieval","late interaction","4096-dim token vectors","layout + table matching"],"url":"https://huggingface.co/tencent/EVIE-8B","notes":"ColPali-style multi-vector retriever on a Qwen3.5 backbone with bidirectional attention, which Tencent reports as rank 1 on ViDoRe V3 at release. Also serves as the teacher for EVIE-4.5B, which supports 64 to 2048 dim prefix embeddings for cheaper indexes. Paper still pending, so scores are self-reported."},{"id":"splade-v3","name":"SPLADE v3","publisher":"Naver Labs","domain":"retrieval","params":"110M","pricing":"Open weights, free to self-host","openWeights":true,"license":"CC-BY-NC-SA-4.0","released":"2024-03","benchmark":"BEIR-13 avg nDCG@10 51.7","capabilities":["sparse retrieval","BM25-compatible inverted index","lexical match expansion"],"url":"https://huggingface.co/naver/splade-v3","notes":"Learned sparse retrieval. Drops into existing inverted-index infra (Elasticsearch, Lucene) without dense vector overhead. Strong hybrid-search complement to dense embeddings. Gated on Hugging Face and non-commercial licensed."},{"id":"esmfold2","name":"ESMFold2","publisher":"Biohub","domain":"science","params":"Folding model on the 6B-parameter ESMC protein language model","pricing":"Open weights, free to self-host; also on Biohub Platform","openWeights":true,"license":"MIT","released":"2026-05","benchmark":"FoldBench: 55% antibody-antigen, 71% protein-protein (77% with MSA)","capabilities":["protein structure prediction","protein + DNA/RNA + ligand complexes","binder design","optional MSA input"],"url":"https://biohub.ai/esm/protein","notes":"Next-generation ESM structure model released by Biohub alongside ESMC and the ESM Atlas. Biohub reports it beats AlphaFold 3 at antibody-antigen pose prediction from single sequences, and used it to design lab-validated binders against five cancer and immunology targets. ESMFold2-Fast folds a 1,024-residue protein in 9.4 seconds."},{"id":"leanstral-1.5","name":"Leanstral 1.5","publisher":"Mistral","domain":"science","params":"119B MoE / 6.5B active","pricing":"Open weights; free API endpoint (labs-leanstral-1-5)","openWeights":true,"license":"Apache-2.0","released":"2026-07","benchmark":"miniF2F 100%, PutnamBench 587/672, FATE-H 87%","capabilities":["Lean 4 theorem proving","autoformalization","code verification agent","256k context"],"url":"https://mistral.ai/news/leanstral-1-5/","notes":"Mistral's formal-math specialist, built on the Mistral Small 4 family. Works as an agent inside a Lean project, reading compiler feedback and editing files across millions of tokens; Mistral puts PutnamBench cost at about $4 per problem. Mistral says it surfaced 5 previously unknown bugs across 57 repositories."}]}