Written by Henk Boelman on October 31, 2026

Mistral Large 4 "Le Chonk" goes open weight: a trillion parameters you can download

Mistral has released the open weights of Mistral Large 4, its trillion-parameter “Le Chonk” model trained in Europe. We look at what it can do, what the license allows and what it takes to run it yourself.

6 min read

Mistral Large 4 - Video

On 29 October, Mistral released the weights of Mistral Large 4, its 1-trillion-parameter flagship nicknamed "Le Chonk." The name started as a joke about a very large cat, but the model is serious: Mistral calls it the best open-weight model for coding, and it was trained from scratch in Europe.

For our community this is the interesting part. Open weights mean you can download the model, inspect it, fine-tune it on your own data and run it on your own infrastructure, without depending on anyone's API. Let's look at what's in the box.

From preview to open weights in three weeks

Mistral launched Large 4 as a public preview on 6 October and promised the weights by the end of the month. It kept that promise, and used the extra time well: three more weeks of reinforcement learning (RL), training the model by rewarding good results, which Mistral says brought big gains in coding and defensive cybersecurity. The released model is therefore better than the one people tried in the preview.

What kind of model is it?

Mixture of experts. Large 4 has about 1 trillion parameters in total, but it's a mixture-of-experts (MoE) model: for each token it only activates a small group of specialized sub-networks, about 49 billion parameters according to Mistral. That keeps it far cheaper to run than its size suggests.
Multimodal and multilingual. It takes both text and images as input and is natively fluent in more than 160 languages.
One model for chat, reasoning and agents. Instead of separate instruct and reasoning models, Large 4 does both. You choose a reasoning mode: Low for simple tasks, Medium for everyday work and High (the default) for the hardest problems. That lets you trade accuracy against cost per request.

Built for coders

Coding is where Mistral makes its biggest claims:

On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with GLM-5.3, the leading open-weight model, while using 47% fewer tokens per task. Tokens are the units models process and bill by, so fewer tokens means cheaper and faster agents.
On Terminal-Bench 4.0, which tests coding in a terminal, it jumped 13 percentage points since the preview to 41%, the top score among open-weight models.
On Cognition's private FrontierCode 2.0 benchmark, which checks whether a maintainer would actually merge the agent's change, it leads open models with 19.7%.
In Surge's blind human preference tests on code, reviewers preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Claude Opus 5.5 came out ahead.

Trusted by defenders, with caveats

Mistral positions Large 4 strongly for defensive cybersecurity: finding vulnerabilities, reproducing them and writing patches. It scores 51% on the Artificial Analysis Cyber Index, and partners like Cisco, CrowdStrike, Cloudflare, Orange Cyberdefense and ENISA, the EU's cybersecurity agency, are testing it.

Mistral is also open about the risks. It says the model refuses 96.6% of malicious cyber requests, the highest rate it measured among open models. But it also reports that monitoring of live traffic showed potential uplift for real attacks, and that it is tightening its guardrails. A specialized version for penetration testing is only available to vetted partners under contract. This is the core tension of open weights: once the weights are public, anyone can modify them, and nobody has yet shown how much of that safety training survives fine-tuning.

Beyond code

Large 4 also improved on broader work tasks. It reaches 73.3% on FinWorkBench (Finch) for finance and spreadsheet workflows, 84.5% on BrowseComp for web research, and resists 97.1% of attacks on the B3 Agent Security Benchmark, which tests things like prompt injection, where hidden instructions try to hijack an agent.

What "open weight" means here

Open weight is not the same as open source. You get the trained model, but not the training data or full training code, although Mistral has promised a technical report. Large 4 is released under the Mistral Ecosystem Open-Weight license, a modified MIT license. You can use, modify, fine-tune and commercialize it freely, with one condition: companies that offer the model as a hosted service (Model-as-a-Service) need a separate commercial license from Mistral. For most builders, startups and enterprises running it for their own products, that's a very permissive deal.

How to try it

API: Mistral Studio, the Mistral API and OpenRouter, priced at $0.70 per million input tokens ($0.07 cached) and $2.10 per million output tokens.
Your own infrastructure: download the weights from Hugging Face and serve them with vLLM or SGLang, two popular open source inference engines. NVIDIA also offers hosted endpoints and an optimized NIM container.
Managed cloud: Microsoft Foundry is the first cloud provider to offer it, with private deployment options on Azure.
In your editor: try it on your own repository in Mistral's Vibe Code.

A reality check on self-hosting: even in compressed formats, the weights take hundreds of gigabytes to around a terabyte of GPU memory. This is a model for a serious multi-GPU server, not your laptop. For most experiments, the API is the easier start.

Our take as a community

Open weights at this level are good news for builders. They give teams real choices: run the model where your data lives, adapt it to your domain and avoid lock-in to a single provider. For organizations in Europe, a frontier-class model trained in Europe with clear origins also matters for sovereignty, which is why names like ABN AMRO, ASML, Crédit Agricole and ESA are already on board.

A few questions to keep in mind:

Check the benchmarks yourself. Most results come from Mistral's own announcement. The weights are public, so independent evaluations will follow quickly.
Mind the hardware. Open weights are only useful if you can run them. Expect smaller, distilled versions to matter as much for most of us.
Safety in the open. The cybersecurity strengths that help defenders could also help attackers. How the community handles fine-tuned versions will be worth watching.

What's next

Mistral says Large 4 will be the foundation for a new generation of specialized models, and a full technical report is coming. We'd love to see what the community builds with it. Running it, fine-tuning it or benchmarking it on your own tasks? Share your results in the community channels or bring them to your next chapter meetup or AgentCon stop.