Mistral Large 4 (Le Chonk): What Developers Need to Know

Mistral Large 4 is a 1-trillion-parameter, natively multimodal model that Mistral AI opened as a public preview on October 6, 2026. You can call it through the API today, but the open weights are promised for the end of the month and aren't downloadable yet.

Mistral nicknames it “Le Chonk.” Every benchmark number below comes from Mistral's own launch post, so treat it as a vendor claim until independent evaluations catch up.

Key Takeaways

• The preview API is live on Mistral Studio. Weights are promised “by the end of the month,” and license terms are still unannounced.
• Per Mistral's docs: 1.05T total parameters, 52B active per token, a 1M-token context window, image input, and 160+ languages.
• Preview pricing is shown as $0.68 input and $2.09 output per million tokens, discounted from $1.36 and $4.18.
• Mistral reports standout cybersecurity results (82% on CyberGym-E2E, 93% on Cybench) and solid but not leading coding scores. All of it is self-reported.
• Open weights don't mean easy self-hosting. A model this size needs serious hardware.

What Is Mistral Large 4?

Mistral Large 4 (ML4) is Mistral's largest model to date: a mixture-of-experts model with 1.05 trillion total parameters and 52 billion active per token, plus a 1.6B vision encoder, according to Mistral's model docs. It takes image input, covers more than 160 languages, and the docs list a 1M-token context window.

It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the preview is served from the same infrastructure. Mistral pitches that as a sovereignty argument: a European deployment it operates end to end, under European law.

One discrepancy to know about: Mistral's launch post and docs say 52B active parameters, while some coverage cites 49B. A secondary report says the Hugging Face placeholder page describes 49B per token, or 52B including embeddings and output layers. This article uses Mistral's own number.

How Much Does the Mistral Large 4 API Cost?

The docs page shows sale pricing of $0.68 per million input tokens, $0.07 per million cached input tokens, and $2.09 per million output tokens. Those are displayed as discounts from $1.36, $0.14, and $4.18. Mistral's own pages don't say how long the sale runs, so check the docs before budgeting around the lower numbers.

Features listed for the model include function calling, structured outputs, document Q&A, batching, and the agents and conversations APIs with built-in tools.

How Do You Call Mistral Large 4?

The docs list the model under the /v1/chat/completions endpoint with the preview ID mistral-large-4. A minimal request follows the standard Mistral chat format:

curl https://api.mistral.ai/v1/chat/completions \
-H "Authorization: Bearer $MISTRAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "mistral-large-4", "messages": [{"role": "user", "content": "Summarize this stack trace."}]}'

We haven't run this against the preview, and preview model IDs can change, so confirm the ID in the docs first.

How Good Is Mistral Large 4 at Coding and Agents?

Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. Its combined Coding Agent Index score is 49.8%, which Mistral says is ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. On AutomationBench, which covers 657 business workflows across apps like Gmail, Google Sheets, Slack, and Salesforce, it scores 59.9%.

The more useful data point may be the blind human evaluation Mistral ran with Surge AI on code quality. Annotators rated ML4 Preview 3.74 out of 5, second of five models, ahead of Kimi K3 (3.59) and GLM-5.3 (3.60) and behind only Claude Opus 5 (4.22). Five models is a small field and Mistral chose the comparison set, so read it as a sanity check rather than a ranking.

Why Is the Cybersecurity Result Getting Attention?

On the Artificial Analysis Cyber Index, Mistral says ML4 ranks in the global top five. On one test in that index, where a model reproduces a real vulnerability in open-source software and then patches it, Mistral reports 82%, the highest of any model. It also reports 93% on Cybench, a set of 40 exercises drawn from security competitions.

Mistral also claims that Claude Opus 5.5 and GPT-6 Astra score near zero on that same test because they refuse the task. That's Mistral's framing, but the underlying point is real for defenders: proving a flaw is exploitable usually comes before fixing it, and a refusal shows up as a failed task, not a capability gap.

Until the weights ship, Mistral says it is red-teaming the model with cybersecurity leaders, vetted partners, and state authorities, who get a version with reduced moderation and expanded cyber capabilities. The public preview is the moderated one. Mistral's own safety section says the model's refusal rate on malicious cyber prompts is higher than every other open-weight model it compared against.

When Do the Weights Drop, and Under What License?

Mistral's wording is “by the end of the month,” with more detail on architecture, benchmarks, and post-training promised alongside. Specific dates differ by outlet: some reports say October 27, while a Hugging Face placeholder page is reported to list October 31. Treat the exact date as unconfirmed.

The license hasn't been announced. Large 3 shipped under Apache 2.0, but at least one report expects a custom license for Large 4. That detail decides whether you can build a product on the weights or need to read field-of-use restrictions first.

Open weights also don't make a 1T-parameter model cheap to run. Only about 52B parameters are active per token, but a mixture-of-experts model still has to hold every expert in memory, so self-hosting means a serious GPU cluster. For most teams the API is the practical way to evaluate it first.

Wrapping Up

Mistral Large 4 is worth a short evaluation on your own tasks, especially if you do security work or need European hosting. Don't plan around the open weights until the license and download actually land. Run a day of real traffic through the preview API while the sale pricing is shown, then revisit when the weights are out.

Sources

• Introducing Mistral Large 4 (Mistral)
• Mistral Large 4 model documentation (Mistral Docs)
• Mistral unveils le Chonk (Tech.eu)
• Mistral's 1T-parameter Large 4 slated for October 31 drop (AI Weekly)
• Mistral Large 4 Preview: What's Confirmed and What's Hype (Online Tech Tips)