Mistral releases the open weights of Large 4, its 1 trillion parameter model

Three weeks ago Mistral promised that its biggest model would not stay locked behind an API. On Thursday, October 29, 2026, the Paris based lab kept that promise. The weights of Mistral Large 4, the model the team calls "Le Chonk", are now available for download on Hugging Face. Anyone with enough hardware can run, inspect and fine tune a model with about 1 trillion parameters, and companies can deploy it inside their own data centers instead of sending data to a cloud provider.

The release is not just the preview checkpoint with a download button. Mistral says it used the weeks since the public preview on October 6 to keep scaling reinforcement learning, and that the version it published is noticeably stronger at coding and security work while needing fewer tokens to reach an answer. That matters for the economics of running agents, where every extra reasoning step costs money.

Here is what was released, what the numbers say, what it takes to run the model and why this download is a bigger deal for Europe than another benchmark table.

What exactly is in the download

According to the model card on Hugging Face, Mistral Large 4 is a natively multimodal mixture of experts model. It has roughly 1 trillion parameters in total, but only 49 billion are active for each token. If you count the embedding and output layers, the active figure is 52 billion, which is why the repository name ends in "A52B". The model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own data centers in Europe.

The key design points are:

• Experts. Each layer has 128 routed experts. For every token the router picks 4 of them, and one shared expert is always used. This is how the model keeps the knowledge of a 1 trillion parameter network while paying the compute cost of a much smaller one.
• Context. The context window is 1 million tokens, large enough for whole code repositories or long contract bundles.
• Input and output. It accepts text and images and answers in text. A Pixtral style vision encoder handles images of different sizes and aspect ratios.
• Attention. The Hugging Face launch blog explains that the 72 layer text decoder combines multi head latent attention with a sparse attention indexer. The indexer selects up to 2,048 earlier positions for each query, a technique also used in GLM 5.3 and DeepSeek models. So the model is sparse twice: it uses only a few experts per token and only a part of the context per query.
• Reasoning modes. One set of weights covers chat, reasoning and agent work. Developers choose low, medium or high reasoning effort per request, with high as the default. Low replaces the old "no reasoning" mode.
• Formats. Mistral ships official FP8 and NVFP4 checkpoints. The community, including Unsloth and the llama.cpp team, published GGUF quantizations on the same day. The checkpoint also includes a built in draft head for speculative decoding, which can speed up generation.
• Languages. More than 160 languages, including every official language of the European Union.

Support arrived on day one. Transformers v5.22.0 added the architecture, and vLLM, SGLang and llama.cpp can serve it.

The benchmark picture

All results below are Mistral's own measurements unless noted, so treat them as claims until independent evaluations arrive. They are still useful, because they show where Mistral thinks the model is strongest.

Coding. On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48 percent. That ties it with GLM 5.3 from Z.ai, until now the leading open weight model on that index, but Mistral says its model needs 47 percent fewer tokens per task on average. On Cognition's private FrontierCode 2.0 benchmark, which judges whether a repository maintainer would actually merge an agent's change, it leads the models Mistral compared with 19.7 percent. The low absolute number shows how hard that test is, not that the model is weak. On Terminal-Bench 4.0, which measures work inside a command line, it reaches 41 percent.

Human preference. In blind side by side ratings on Surge Code, graders preferred Mistral Large 4 over GLM 5.3, Kimi K3 and DeepSeek V4.1 Flash. Among the models evaluated, only Claude Opus 5.5 ranked higher.

Security. Mistral reports 51 percent on the Artificial Analysis Cyber Index, which tests whether a model can find a vulnerability in a codebase, reproduce it and ship a patch without breaking other features. At the same time it says the model refuses malicious cyber requests 96.6 percent of the time on average across JailbreakBench, StrongREJECT and AgentHarm, the highest rate it measured among open models.

Office work and browsing. On FinWorkBench, also called Finch, which focuses on spreadsheet heavy finance and accounting tasks, it scores 73.3 percent, second among open models after Kimi K3. On BrowseComp, a test of finding hard to locate facts on the web, it reaches 84.5 percent, up from 75 percent for the supervised checkpoint the reinforcement learning started from.

Hugging Face adds an outside reference point: as of October 29, the model ranks fifth among open weight models on the Artificial Analysis Intelligence Index. That is a useful reality check. Le Chonk is very strong in the areas Mistral optimized for, but it is not the best open model at everything.

Three weeks of reinforcement learning made a difference

The most interesting part of this release is the comparison with the preview. Mistral said in early October that its reinforcement learning run was still going and that it expected "large and rapid improvements". The published numbers back that up, at least on the tests Mistral chose to repeat.

On Terminal-Bench 4.0 the score climbed 13 percentage points, from 28.3 percent in the preview to 41 percent, while output tokens fell by 30 percent. On the B3 agent security benchmark, which tests whether an agent can be tricked into leaking its system prompt or running injected code, attack resistance went from 93.3 to 97.1 percent. For companies that want to put an agent in front of customers or connect it to internal tools, that second number may matter more than any coding score.

Mistral also writes that its gains "continue to accelerate, with no signs of saturation". That is marketing language, but it hints that Large 4 is meant as a base for more releases. The company says it will build a new generation of specialized models on top of it and that a full technical report is coming soon.

Built for defenders, with guardrails that will change

Mistral leans hard on cybersecurity, and it explained its thinking to TechCrunch. Pierre Stock, Mistral's VP of Science, said open weights give businesses and governments "access to frontier intelligence to defend yourselves against cyberattacks. Access that can't be cut off by definition because the model is open-weighted." His argument is that attackers will use AI either way, so defenders should get the same capability as early as possible.

The launch post lists quotes from Booz Allen, Cisco, Cloudflare, Fortinet, Rubrik and Zscaler, plus banks such as ABN AMRO, Belfius and La Banque Postale. Booz Allen says the model matched the vulnerability detection recall of other frontier cyber models on its internal source code benchmark at an attractive cost.

Mistral is also open about the risk. It says that during the preview it monitored cybersecurity related use and saw the model refuse black box attacks on live targets, but it also saw "a potential for uplift to conduct real attacks". It is adjusting the guardrails and plans more updates in the coming months. A separate, more permissive variant for penetration testing and red teaming is available only to vetted partners such as Orange Cyberdefense, under contract and only for systems they own or are allowed to test.

There is a tension here that every open model maker faces. Once weights are public, the guardrails Mistral trains into them can be weakened by anyone who fine tunes the model. Mistral's answer is that refusals are part of the released weights and that broad access helps defenders more than attackers. Whether that holds will only become clear after independent red teams have worked with the model for a while.

The license: open, with one commercial catch

Mistral Large 4 ships under the Mistral Ecosystem Open-Weight License, a modified MIT license. You can use, modify, fine tune and commercialize the model for free. The single condition: a company that runs a model as a service business, meaning it sells hosted access to the model itself, needs a separate commercial license from Mistral first.

For most companies that simply want to build an internal assistant, a coding agent or a product on top of the model, this changes nothing. For cloud hosts and inference startups that would otherwise undercut Mistral's own API with the same weights, it does. It is a middle path between a fully permissive license and the stricter licenses some other labs use.

What it takes to run Le Chonk

Open weights do not mean you can run this model on a laptop. A rough rule of thumb is one byte per parameter in FP8, so the weights alone take about 1 TB of memory. The NVFP4 checkpoint brings that down to roughly 0.55 TB. On top of that you need memory for the context cache, which grows with long prompts.

In practice that means a full GPU server, such as eight Nvidia H200 cards with 141 GB each, or a Blackwell system for the NVFP4 version. Even Apple's new Mac Studio with 512 GB of unified memory falls short for the official checkpoints, although heavily quantized GGUF files from the community will push the limits for hobbyists. These are our own estimates, not Mistral figures.

For everyone else there are hosted options. The Mistral API charges $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The model is also on OpenRouter, Microsoft Foundry, which Microsoft says makes it the first cloud provider to offer it, and NVIDIA's hosted endpoints and NIM containers. Hugging Face offers Inference Endpoints and serverless access through its Inference Providers.

Why this matters

For two years, the largest open weight models mostly came from China, while the strongest American models stayed closed. Mistral Large 4 is an open model at the 1 trillion parameter scale that was trained from scratch in Europe, on European infrastructure, by a European company. For banks, public agencies and defense contractors that worry about where their AI comes from and who can switch it off, that is the main selling point, and Mistral says so openly when it talks about "sovereign AI".

For developers, the practical news is simpler. There is now another top tier open model for coding agents that uses far fewer tokens than its closest competitor, with day one support in the usual tools and a license that allows commercial use. If Mistral's efficiency claims hold up in independent tests, running agents on it could cost noticeably less than on GLM 5.3 at a similar quality level.

The open questions are the usual ones. Independent benchmarks still have to confirm Mistral's numbers. The technical report is not out yet. And the security guardrails will be tested by people with very different goals than Mistral's design partners. For now, the promise from October 6 has been kept, and the model is a download away.

Sources: Mistral AI announcement, Hugging Face model card, Hugging Face blog, Transformers v5.22.0 release notes, TechCrunch, Mistral docs.