Mistral AI on October 29, 2026 released open weights for Mistral Large 4 (nickname Le Chonk), its roughly 1-trillion-parameter natively multimodal Mixture-of-Experts model, after a public preview earlier in the month and further reinforcement learning focused on coding and defensive cybersecurity. Official weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B (FP8) and an NVFP4 variant, with day-0 serving recipes from vLLM and GGUF builds from Unsloth and llama.cpp.
Key Facts
| Item | Detail |
|---|---|
| Model | Mistral Large 4 (Le Chonk) |
| Scale | ~1T total parameters; 49B active per token (52B including embeddings and output layers), per the official Hugging Face model card and vLLM |
| Type | MoE instruct/reasoning multimodal (text + image in, text out); Pixtral-style vision encoder |
| Context | 1M tokens (model card) |
| MoE layout | 128 routed experts, top-4 active + 1 shared expert; 72 decoder layers (layer 0 dense) |
| Reasoning modes | low / medium / high (default high) |
| Weights | Hugging Face FP8 and NVFP4; Unsloth and llama.cpp GGUF quants |
| Serving | vLLM and SGLang day-0; also Mistral API/Studio, OpenRouter, Microsoft Foundry, NVIDIA NIM |
| License | Mistral Ecosystem Open-Weight License (MIT-derived); free to use, modify, fine-tune, and commercialize, except Model-as-a-Service needs a separate commercial license from Mistral |
| API price (Mistral) | $0.70 / M input, $0.07 / M cached input, $2.10 / M output |
| Training claim | Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral datacenters in Europe (model card) |
Selected scores Mistral cites for the open-weight release (company blog unless noted):
- Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3, with ~47% fewer tokens per task on average
- Terminal-Bench 4.0: 41% (+13 percentage points vs public preview; ~30% fewer output tokens)
- Cognition FrontierCode 2.0: 19.7% (Mistral says it leads open weights; Claude Opus 5.5 is far higher in THE DECODER's report at 46%)
- Artificial Analysis Cyber Index: 51% (Mistral frames open-weight leadership; THE DECODER notes Mistral's own chart still shows DeepSeek V4.1 Flash and MiMo-V2.6-Pro ahead among open models)
- Malicious cyber refusal: average 96.6% on cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm (Mistral)
- Artificial Analysis Intelligence Index v5: 26 (model card; THE DECODER matches; highest among Western open weights on that board, behind several Chinese open models)
Hugging Face's companion blog lists some overlapping metrics with different numbers (for example Coding Agent Index 49.8%, Terminal-Bench 4 28.3%, Lakera B3 attack resistance 93.3%). Treat those as HF's published snapshot; Mistral's open-weight post and model card are the primary figures above for Coding Agent Index, Terminal-Bench gains, and B3 97.1% vs preview 93.3%.
Background
Mistral had already shipped Large 4 in public preview earlier in October 2026. The open-weight cut follows additional RL that the company says improved coding agents, defensive cyber workflows, token efficiency, and several enterprise evals (Finch spreadsheet work to 73.3%, BrowseComp to 84.5%). Architecture write-ups from Hugging Face and vLLM describe Multi-head Latent Attention with a DeepSeek-style sparse indexer (top-2048 keys), cross-layer KV sharing, NoPE layers, and an embedded DSpark speculative-decoding drafter in the checkpoint.
Availability spans self-host and managed paths: download weights on Hugging Face; serve with vLLM (including NVFP4 on Blackwell) or SGLang; use Mistral's API and Studio; or run through partners such as Microsoft Foundry and NVIDIA-hosted/NIM options. Unsloth published GGUF packages that include the vision projector and DSpark draft weights for local llama.cpp builds.
Reactions and Impact
Partner quotes on Mistral's announcement emphasize self-hosting and data residency for security and banking workloads (for example Booz Allen on source-code security analysis, Cloudflare on auditing and on-prem control, and several European banks on sovereign deployment). Hugging Face product lead Jeff Boudier framed the release as giving builders open foundations for coding, multimodal use, and cyber defense without depending on a single hosted API.
Independent coverage from THE DECODER stresses coding strength and lower API list prices versus the preview, while arguing the model still trails leading Chinese open systems on Artificial Analysis's overall Intelligence Index and that Cyber Index leadership is more contested than Mistral's marketing copy implies. THE DECODER also reports Artificial Analysis throughput around 161 tokens/s for the final high-reasoning setting and a lower average cost per index task than the preview.
What We Don't Know Yet
- How third-party leaderboards will settle Coding Agent Index / Terminal-Bench numbers where Mistral and Hugging Face posts already disagree by a point or more
- Whether Artificial Analysis will keep ranking Large 4 behind other open models on Cyber Index after more independent runs
- Hardware floor and cost for comfortable self-host at full precision beyond recipe examples (for example multi-GPU B300/GB300 setups in partner docs)
- Timeline and scope for smaller or specialized variants Mistral says will build on Large 4
- Commercial MaaS license terms and pricing beyond the public "separate license required" clause