On October 29, 2026, Mistral AI released the weights of Mistral Large 4, nicknamed "Le Chonk." It is a 1-trillion-parameter, natively multimodal model that first appeared in public preview earlier this month. Anyone can now download it from Hugging Face and run it on their own hardware, but one group of businesses needs a separate license first.

Coffee summary

  • FACT: Mistral Large 4 is a 1-trillion-parameter multimodal model, and its weights are now downloadable from Hugging Face.
  • FACT: It ships under the Mistral Ecosystem Open-Weight license, a modified MIT license. Companies that sell the model as a hosted service need a separate commercial license.
  • FACT: API pricing is $0.70 per million input tokens, $0.07 cached input, and $2.10 output.
  • CLAIM: Mistral says it is the best open-weight model for coding, citing a 19.7% score on Cognition's private FrontierCode 2.0 benchmark.
  • OPINION: For teams worried about depending on one closed API, this is the most serious open fallback model so far this year.

What happened

Mistral says three extra weeks of reinforcement learning after the preview produced big gains in coding and defensive cybersecurity, and that the model now needs fewer tokens to reach an answer. The release adds three reasoning modes: Low for simple tasks, Medium for everyday work, and High (the default) for the hardest problems.

You can use it in Mistral Studio, through the Mistral API, on OpenRouter, through NVIDIA-hosted endpoints, or in Microsoft Foundry. You can also download the weights and serve them yourself with vLLM or SGLang.

The numbers Mistral is leading with

Benchmark (Mistral's figures) Score
Artificial Analysis Coding Agent Index 48% (Mistral says it is roughly tied with GLM-5.3 while using 47% fewer tokens)
Terminal-Bench 4.0 41%, up 13 points from the preview
Artificial Analysis Cyber Index 51%
Refusal of malicious cyber requests 96.6% average
BrowseComp (web browsing) 84.5%

CLAIM: These are vendor-selected results. Independent coverage is more cautious. DeAI's daily brief, for example, reports that Artificial Analysis scores it at 26 points, behind DeepSeek V4.1 Flash and GLM-5.3 at 32 each. Treat the chart as a shortlist signal, not a final ranking.

Why it matters

Open weights change who holds control. With a closed model, the vendor decides on price changes, rate limits, and retirements. With downloadable weights, you can pin a version, fine-tune it on private data, and run it inside your own network. Mistral pitches this as "sovereign AI," and banks and security vendors quoted in the launch post say data residency is the main draw.

The security angle matters too. Mistral says the model handles defensive work such as finding, reproducing, and patching vulnerabilities without needless refusals. It also admits that monitoring showed "a potential for uplift to conduct real attacks," and says it is tightening guardrails. A more specialized cyber variant is limited to vetted partners.

Who should care

  • Developers and startups: anyone who wants a coding-capable model they can host, or a backup if an API price changes.
  • Regulated industries: banks, governments, and healthcare teams with strict data-residency rules.
  • Security teams: people testing AI for vulnerability triage on systems they own.

Limitations

  • Hardware: a trillion-parameter model is not something you run on a laptop. Self-hosting needs serious GPU capacity.
  • License: "open weight" is not the same as "anything goes." If you plan to resell access as a hosted model, read the license file and talk to Mistral first.
  • Docs: Mistral says a full technical report is "coming soon," so independent verification is still limited.
  • Gap to the top: in the Surge Code human-preference results Mistral published, it still ranks behind Claude Opus 5.5.

What to do next

  1. Add Large 4 to your internal evals next to the model you pay for now, using your real tasks.
  2. Start on the API ($2.10 per million output tokens) before buying hardware.
  3. If self-hosting, check the model card for supported precisions and serving requirements.

AIImpish Take

Mistral Large 4 is the strongest sign yet that frontier-scale models do not have to stay locked behind one company's API. It will not beat the very best closed models on every task, but owning a capable fallback is valuable on its own. Builders who depend on a single provider should test it this week; everyone else can wait for independent benchmarks.

*Sources: Mistral, "Mistral Large 4 (Le Chonk) is now open weight"; Hugging Face blog, "Welcome Le Chonk"; TechCrunch, Oct 29, 2026; DeAI Daily Brief, Oct 30, 2026.*