Mistral Large 4 is a trillion-parameter mixture-of-experts model from the European AI company Mistral, nicknamed Le Chonk, with roughly 49 billion parameters active per task. A preview is live now, and downloadable open weights were slated for the end of October 2026. Most capability claims below are Mistral-reported and still unverified by independent testing.
What Is Mistral Large 4?
Mistral Large 4 is Mistral's largest and most capable model to date, announced in October 2026 under the nickname Le Chonk, a fat-cat joke about its size. Mistral is a European AI company, and it states the model was trained from scratch in its own European data centers. A preview is currently available online, with downloadable open weights planned for the end of October 2026.
The model is natively multimodal, meaning it processes images as well as text. Mistral also reports that it trained on more than 160 languages, including every official language of the European Union. You can try the preview through Mistral's online studio, and the company's announcements live on mistral.ai.
One caveat belongs up front: the video this article is based on was published one day after launch, and its speaker explicitly had not run independent tests. Every benchmark and capability figure below is Mistral-reported unless stated otherwise.
How the Trillion-Parameter Sparse Architecture Works
Le Chonk does not fire all one trillion parameters at once. For any single task, about 49 billion parameters activate, which is what a mixture-of-experts (MoE) design does: a routing system picks the relevant expert subsets for each input, while the rest stay idle.
The practical effect is a model with very large total capacity but per-task compute closer to a much smaller model. Mistral also frames the current release as unfinished: the company says training is still ongoing and that the preview will keep improving. It describes the model as a foundation for a planned generation of task- and industry-tuned variants rather than a final product.
Treating launch numbers as provisional matters here. If Mistral continues training through the open-weights release, benchmark results published at preview time may not describe the weights people actually download.
Open Weights vs the Live Preview: What Ships When
There are two ways to use this model, and they arrive at different times. The preview is available now through Mistral's hosted interface. The open-weights version, which you can download and run on your own hardware under your own rules, was scheduled for the end of October 2026.
The distinction matters for different audiences. Security teams, researchers and enterprises that cannot send data to a hosted endpoint care mainly about the downloadable weights. Anyone evaluating capability today can only judge the preview, which Mistral itself describes as an early, still-improving version.
A practical sequence for evaluating it: try the preview on representative tasks; wait for the open-weights release on Mistral's channels; check the weights and model card on Mistral's Hugging Face page when they land; and re-run your own tests, because the downloaded checkpoint may differ from the preview.
Reported Capabilities: Coding, Agents, Vision and Security
Mistral positions Le Chonk as one model covering instruction following, step-by-step reasoning, agentic work and vision, instead of separate specialized models. All results below are vendor-reported from launch materials in October 2026.
On coding, Mistral says the model scores ahead of other large open models such as DeepSeek and Qwen on agentic coding benchmarks. In a blind human-preference test, it reportedly ranked second out of five, behind Claude. On business workflows, Mistral reports it came out ahead of other top open models across a 657-workflow test spanning applications like Gmail, Slack and Google Sheets, and it handles spreadsheet building, slide creation and document drafting.
Vision is the less-discussed capability. Mistral demonstrated the model zooming into an engineering drawing to verify a part, and scanning a large satellite image to spot small ground features, a use case relevant to disaster response. The company reports it edged out a top closed model on one vision benchmark. Mistral also says it leads other open models on tough scientific coding tasks and can write a full chemistry simulation in one pass.
For security teams, Mistral claims it is one of the strongest models for defensive cybersecurity: finding software weak spots so teams can patch them. Running the open weights on local infrastructure lets a security operation keep the model working under its own rules. Mistral additionally reports the model resisted more than 9 out of 10 attempts to push it off safe behavior in testing.
A summary of who reported what:
Capability Claims and Their Evidence Status
| Capability | Reported result | Evidence source |
|---|---|---|
| Agentic coding | Ahead of DeepSeek and Qwen | Mistral-reported, unverified |
| Blind human preference | 2nd of 5, behind Claude | Mistral-reported, unverified |
| Business workflows | Ahead of top open models across 657 workflows | Mistral-reported, unverified |
| Vision | Edged out one top closed model on one benchmark | Mistral-reported, unverified |
| Robustness | Resisted >9/10 adversarial attempts | Mistral-reported, unverified |
Read the table as a to-verify list, not a scoreboard. Vendor benchmark tables prove their own setups, and none of these figures has an independent reproduction attached to it as of October 2026. When the open weights ship, third parties can rerun the standard suites, and that is when the claims become checkable.
Multilingual Training: 160+ Languages and EU Coverage
Mistral states the model was trained on more than 160 languages, covering every official language of the European Union. For a European lab, that is a positioning choice: most flagship open models come from the US or China, and broad EU language coverage is a differentiator Mistral is emphasizing.
The practical implication for users is narrower than the headline. Training-language counts do not tell you per-language quality, and Mistral has not published per-language benchmark breakdowns in the launch material referenced here. If you work in a specific EU language, test the preview in that language before committing a workflow to it.
Should You Use Le Chonk, and For Which Jobs?
Match the model to the job. Le Chonk is a heavy-duty model aimed at agentic coding, multi-app workflows, document and image understanding, and defensive security work. For small, simple tasks, a lighter model is sufficient and cheaper, and spending a trillion-parameter model on trivial work wastes resources.
If you do adopt it, the vision side is the capability most likely to be missed. Users who treat it as a plain chatbot ignore the image and document understanding that Mistral presents as a key feature. Feeding it charts, scans and drawings is where the vendor claims the strongest surprises.
Timing also matters. The current preview is explicitly early and still improving, so results you get this week may undersell the final release. For anything production-critical, waiting for the open-weights release at the end of October 2026 and running your own evaluation is the defensible move.
Frequently Asked Questions
- What is Le Chonk? Le Chonk is the nickname for Mistral Large 4, a 1-trillion-parameter mixture-of-experts model announced by Mistral in October 2026. About 49 billion parameters activate per task, and Mistral plans open-weights release by the end of that month.
- Is Mistral Large 4 open source? Mistral has announced open weights, scheduled for the end of October 2026. Open weights let you download and run the model yourself; the hosted preview available now is a separate, earlier access path.
- Does Mistral Large 4 really beat DeepSeek and Qwen? Mistral reports that it leads other large open models on agentic coding and a 657-workflow business test. These are vendor-reported results from launch materials as of October 2026, with no independent reproduction attached yet.
- Can Mistral Large 4 read images? Yes. It is natively multimodal, and Mistral demonstrated it reading engineering drawings, charts and satellite imagery. The company reports it outperformed one top closed model on a single vision benchmark.
- How can I try it today? The preview is live through Mistral's hosted online studio. The downloadable version arrives later, and Mistral cautions the preview is still improving, so evaluate it knowing the final weights may perform better.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits