On Oct. 29, 2026, Mistral released the open weights of an improved Mistral Large 4 (“Le Chonk”), its flagship 1-trillion-parameter, natively multimodal model. The drop follows a public preview earlier in October and lands alongside day-0 support in Hugging Face Transformers, vLLM, SGLang, and llama.cpp, with checkpoints on Hugging Face in FP8 and NVFP4.
Key facts
| Item | Detail |
|---|---|
| Model | Mistral Large 4 (Le Chonk) — 1T-parameter natively multimodal MoE (Mistral) |
| Active params | ~49B per token routed (~52B including embeddings and output layers, per Hugging Face Transformers v5.22.0 release notes) |
| MoE layout | 128 routed experts, 4 active per token, plus 1 shared expert (Transformers release notes) |
| Context & I/O | 1M-token context; text and image input, text output; 160+ languages (Transformers release notes) |
| Reasoning | Three efforts — low, medium, high (default) (Mistral) |
| Coding index | 48% on Artificial Analysis Coding Agent Index, “effectively tied” with GLM-5.3 while using 47% fewer tokens per task on average (Mistral) |
| API pricing | $0.70 input / M, $0.07 cached input / M, $2.10 output / M tokens (Mistral) |
| Licence | Mistral Ecosystem Open-Weight License — modified MIT; separate commercial licence for Model-as-a-Service (Mistral) |
What Mistral released
In Built for Coders, Trusted by Defenders., Mistral says it is releasing “the weights of this improved Mistral Large 4” after scaling reinforcement learning on coding and defensive cybersecurity. Weights are on Hugging Face; developers can try the model in Mistral Studio, the Mistral API, OpenRouter, Vibe Code, and through NVIDIA NIM and hosted endpoints, or self-host with vLLM or SGLang. Mistral notes a full technical report is “coming soon.”
The Hugging Face blog Welcome Le Chonk: Mistral Large 4 on Hugging Face lists official FP8 and NVFP4 variants plus community GGUF builds from Unsloth and the llama.cpp team, all under the same open-weight licence family.
What changed since the preview
Mistral frames the open drop as a refinement pass, not a rename: additional reinforcement learning improved agentic coding and defensive cyber workflows, with “substantial gains in reasoning efficiency” that lower inference cost. The model now exposes three reasoning modes—low for simple tasks (replacing a prior “None” mode), medium for everyday work, and high as the default for the hardest cases.
On Terminal-Bench 4.0, Mistral reports a 13-point gain to 41% versus preview, leading open-weight models while cutting output tokens by 30%. On the Artificial Analysis Cyber Index, it scores 51% on vulnerability discovery, reproduction, and patching. Mistral also cites 96.6% refusals on malicious cyber prompts (JailbreakBench, StrongREJECT, AgentHarm) and 97.1% B3 attack resistance versus 93.3% in preview.
Architecture and specs
Transformers v5.22.0 describes a 1T natively multimodal MoE with 49B active parameters per token (52B with embeddings and output layers), 128 routed experts (4 active per token) plus 1 shared expert, 1M context, Pixtral-style vision, and unified instruct, reasoning, and agentic tools in one checkpoint. Hugging Face's blog adds 72 decoder layers and a 1.6B Pixtral encoder with sparse attention indexing—details Mistral's news post does not spell out.
How it scores
Mistral's own charts anchor the coding story. On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48%, “effectively tied” with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average. On Cognition's private FrontierCode 2.0 benchmark—whether a maintainer would merge an agent's change—Mistral reports a lead at 19.7%.
Hugging Face's launch post lists a 49.8% Coding Agent Index score and separate cyber benchmarks (including 82% on vulnerability reproduction and patching in its Cyber Index discussion)—figures attributed to Hugging Face, not Mistral's news charts.
Where to get it and pricing
Checkpoints live on Hugging Face (FP8 and NVFP4); run locally with Transformers, vLLM, SGLang, or llama.cpp, or use Mistral Studio, the Mistral API, OpenRouter, NVIDIA NIM, Vibe Code, and Forge for hosted or tailored deployments. Mistral prices the API at $0.70 input / M, $0.07 cached input / M, and $2.10 output / M tokens.
The licence catch for Model-as-a-Service
Mistral releases the weights under the Mistral Ecosystem Open-Weight License, a modified MIT licence: free to use, modify, fine-tune, and commercialize, with one carve-out. Businesses that operate a Model-as-a-Service offering must obtain a separate commercial licence from Mistral before commercial use. Hugging Face's Transformers release notes repeat the same MaaS condition.
Why it matters
Analysis. A trillion-parameter European MoE with public weights is a direct bet on sovereign deployment—inspectable checkpoints, on-prem inference, and Forge-style customization without routing sensitive code through a single vendor API. The coding and cyber numbers Mistral publishes are as much about token economics as raw leaderboard rank: tying GLM-5.3 on the Coding Agent Index while cutting average tokens per task matters when agents loop for hours.
The release also stacks on Mistral's broader open-weight line, including voice agents in Mistral Voxtral 3 Brings Voice Agents in 50+ Languages, and sits in the same trillion-class open tier as rivals such as Moonshot's Kimi K3.1 Pro Is a 3T-Class Open-Weight Model—but with Mistral's emphasis on defensive security workflows and EU training lineage.
What to watch
Mistral promises a full technical report; until it lands, architecture claims will stay split between marketing posts and Hugging Face's deep dive. Track whether preview customers move from API-only trials to self-hosted FP8 or NVFP4 clusters, how reasoning-effort defaults shift cost on long agent runs, and whether Mistral's updated cyber guardrails—after live-traffic monitoring flagged potential uplift for real attacks—hold up in independent red-team work.