Image: Mistral AI logo (Wikimedia Commons) — the Paris lab behind Large 4. This is the company's mark, shown for identification; it is not the model itself.
TL;DR — Mistral Large 4 open weights explained in one line: France's Mistral AI has published the full downloadable weights of Large 4 — nicknamed "Le Chonk" — a 1.05-trillion-parameter, natively multimodal model built for agentic coding and defensive cybersecurity, free to run on your own infrastructure under a modified MIT licence, with one catch for model-as-a-service businesses.
📌 KEY HIGHLIGHTS
- Weights out on 29 October 2026: Large 4 moved from an API-only public preview (6 October) to fully downloadable weights on Hugging Face, with no access gate.
- Trillion-parameter scale, sparse cost: 1.05 trillion parameters in total, but only about 52 billion active per token, plus a 1-million-token context window.
- Coding and cyber defence pitch: 41% on Terminal-Bench 4.0 (up from 28% at preview), 48% on the Artificial Analysis Coding Agent Index, and 51% on its Cyber Index.
- Licence with a boundary: the Mistral Ecosystem Open-Weight License allows use, modification and commercialisation — but businesses offering the model as a service need a separate commercial licence.
- Honest caveat: Mistral itself says live-traffic monitoring showed the model could potentially help real attacks, and it is revising its guardrails.
Mistral Large 4 open weights explained: what happened?
On 29 October 2026, Mistral AI released the weights of Mistral Large 4, the flagship model it previewed on 6 October under the nickname "Le Chonk". Reuters reported the public release the same day, with Mistral pitching the model to organisations that want to find and fix cybersecurity vulnerabilities on their own systems rather than through a third-party provider. The weights are published on Hugging Face as mistralai/Mistral-Large-4-1T-A52B, and the company says the model can now be downloaded, run, inspected and fine-tuned independently.
The release matters because it closes the gap between preview and ownership. During the preview, Large 4 was available only through Mistral's API and Studio. With the weights out, a security team can run the model inside its own perimeter — scanning proprietary code and incident data without sending it to an outside provider. Mistral's vice president of science, Pierre Stock, put the philosophy bluntly: "I'm not comfortable living in a world in which open weights do not exist."
The specs that matter
Large 4 is a mixture-of-experts model: instead of firing every parameter for each request, a router activates only a small subset of specialist subnetworks per token. That is why a trillion-parameter model can cost roughly as much per token as a dense model a twentieth of its size — while memory and serving hardware still scale with the full checkpoint.
| Spec | Mistral Large 4 |
|---|---|
| Total parameters | 1.05 trillion (mixture of experts) |
| Active parameters per token | About 52 billion |
| Modalities | Text and image input, text output; 1.6-billion-parameter vision encoder |
| Context window | 1 million tokens |
| Languages | 160+, including every official EU language |
| Training | From scratch on about 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centres |
| Reasoning modes | Low, Medium and High (High is the default) |
| Licence | Mistral Ecosystem Open-Weight License (modified MIT; model-as-a-service needs a separate commercial licence) |
| API price at release | $0.70 per million input tokens, $0.07 cached input, $2.10 output — down from the preview's $1.36 / $4.18 |
| Self-hosting | vLLM, SGLang, Transformers and llama.cpp; NVFP4 quantised checkpoint targets Nvidia Blackwell GPUs |
.jpg?width=900)
Illustrative image: server racks in a data centre (Wikimedia Commons, CC BY 2.0, Tony Webster). Running a trillion-parameter checkpoint takes this kind of infrastructure — it is not a laptop download.
Benchmarks: where it leads, and where it doesn't
Mistral's headline numbers are strong, but they need the right labels: most are company-reported, while the Artificial Analysis Intelligence Index is an independent evaluation. On coding, Large 4 scored 19.7% on FrontierCode 2.0 and 41% on Terminal-Bench 4.0, up 13 points from the preview, with output tokens down 30%. On the Artificial Analysis Coding Agent Index it reached 48% — effectively tied with China's GLM-5.3 while using 47% fewer tokens per task — and blind comparisons by Surge Code ranked it behind only Claude Opus 5.5 among the models tested.
On cybersecurity, Mistral reports 51% on the Artificial Analysis Cyber Index, which tests finding, reproducing, validating and patching vulnerabilities without breaking existing functions, and says the model refuses an average of 96.6% of malicious cyber requests — the highest rate it measured among open models. Independent evaluator Artificial Analysis gives Large 4 at its high reasoning setting 26 points on its Intelligence Index v5, making it the highest-scoring model with available weights developed outside China — but still six points behind the open-weight leaders DeepSeek V4.1 Flash and GLM-5.3, and with cost per task down 63% from the preview ($2.06 versus $5.54). The honest read: best Western open-weight option, not the overall leader.
Why it matters for Europe readers
Large 4 is Europe's sovereignty argument in model form. Trained from scratch on European infrastructure, offered as downloadable weights, and framed for governments and enterprises that cannot depend on a foreign provider's filters or access decisions, it is designed for organisations that treat AI control as a security requirement. Mistral is already testing a specialised cyber variant with vetted enterprise partners — Orange Cyberdefense is the named design partner — for penetration testing, vulnerability assessment and attack simulation on owned or authorised systems.
There is also a warning inside the announcement that deserves equal weight. Mistral says that since the preview, its live-traffic monitoring showed the model could potentially help facilitate real attacks even though it refused direct attacks on real targets, and it is modifying its guardrails in response. Open weights cut both ways: once released, access cannot be contained, as Stock himself acknowledged. Defenders get a powerful tool they fully control; so does everyone else. That is why Mistral's own framing pairs the release with vetting, contractual safeguards and human review — and why the week's other frontier-model news, our Google Gemini 4 Carbon India explainer, shows the opposite strategy: a powerful cyber model kept behind staged, trusted-partner access instead of a public download.
Should you care? ✅ / ⏳
Verdict: ✅ Yes if you build or defend software, ⏳ otherwise — for security and platform teams, a downloadable frontier-scale coding and cyber-defence model with a 1-million-token context is a genuine new option, especially inside regulated or sovereign environments. For everyone else, this is a story to watch, not a download to run: a trillion-parameter checkpoint needs cluster-grade hardware, and the benchmarks that matter most will come from independent teams running the released weights.
Availability & alternatives
The weights are live now on Hugging Face (mistralai/Mistral-Large-4-1T-A52B) with no access gate, and hosted access continues through Mistral Studio, the Mistral API, OpenRouter, Nvidia-hosted endpoints and Microsoft Foundry. If you only need the model through an API, the release-day rates are $0.70 per million input tokens and $2.10 per million output tokens. Alternatives in the open-weight race include Z.ai's GLM-5.3 and DeepSeek V4.1 Flash, which still lead on the broad independent index, and Meta's Muse Spark Open — while India's own sovereign-AI push, covered in our Sarvam AI defence ministry agreement post, shows the same control-versus-capability debate playing out closer to home.
FAQs
What is Mistral Large 4 (Le Chonk)?
Mistral Large 4 is Mistral AI's flagship open-weight model, nicknamed "Le Chonk". It is a natively multimodal mixture-of-experts system with 1.05 trillion total parameters, about 52 billion active per token, a 1-million-token context window and support for 160+ languages.
Where can I download Mistral Large 4 weights?
The weights were published on Hugging Face on 29 October 2026 as mistralai/Mistral-Large-4-1T-A52B, with no access gate. They can be served with vLLM, SGLang, Transformers or llama.cpp, and a quantised NVFP4 checkpoint targets Nvidia Blackwell GPUs.
Is Mistral Large 4 really open source?
Not quite — it is open-weight, not open source in the strict sense. The Mistral Ecosystem Open-Weight License, a modified MIT licence, allows use, modification, fine-tuning and commercialisation, but businesses that operate a model-as-a-service offering must obtain a separate commercial licence from Mistral first.
How much does the Mistral Large 4 API cost?
At release, Mistral lists $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens — substantially lower than the preview's $1.36 input and $4.18 output rates.
Takeaway: Mistral Large 4's open weights give Europe — and any team that needs AI inside its own walls — a frontier-scale coding and cyber-defence model it can actually own, with a licence catch and a misuse warning that deserve as much attention as the benchmarks.
Follow UK & Europe Tech on TechByCountry for tomorrow's update.
📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.
Sources
- Reuters via Northland News Radio — Mistral AI publicly releases Large 4, targets cybersecurity users (29 Oct 2026)
- VentureBeat — Mistral Large 4 'Le Chonk' arrives as the top open model outside China (29 Oct 2026)
- Unite.AI — Mistral's 1-trillion-parameter Le Chonk checkpoint lands on Hugging Face (29 Oct 2026)
- Testing Catalog — Mistral opens Large 4 weights for coding and security (29 Oct 2026)
- AlphaSignal — Mistral drops Large 4, a 1 trillion parameter open coding agent (29 Oct 2026)
Comments
Post a Comment