Mistral Large 4 Open Weights Are Out: 1 Trillion Parameters You Can Download

The Mistral Large 4 open weights went public on October 29, 2026, so anyone can now download the 1-trillion-parameter model from Hugging Face. It is built for coding agents and cybersecurity work. The catch: you need serious GPUs to run it yourself.

Here is what is confirmed so far, and what Mistral hasn’t told us yet.

• Released: weights went live October 29, 2026, after a preview about three weeks earlier
• Size: 1 trillion parameters total, 52 billion active per task
• Where: Hugging Face, Mistral’s API, and some Vibe users
• Hardware: a quantised version runs on four Nvidia B200 or B300 GPUs
• Not yet public: API pricing and license terms (per the reports we checked)

What Mistral Large 4 actually is

Mistral Large 4, nicknamed “Le Chonk,” is a mixture-of-experts style model. That means only 52 billion of its 1 trillion parameters work on any one task, according to The Next Web. It reads text and images and writes text and code.

Mistral first showed it earlier in October, when we covered the Mistral Large 4 preview before the weights were available. It then spent about three more weeks on reinforcement learning, guided by customer feedback, before releasing the weights.

“Open weights” is not the same as fully open source. You can download and run the model, but the training data and code are not part of the deal. Check the license on Hugging Face before using it commercially, since the coverage we found doesn’t state it.

How Mistral Large 4 performs on benchmarks

Mistral says the model beats other open models on agentic coding. These numbers come from Mistral’s own claims as reported by The Next Web, so treat them as a starting point, not a verdict.

The token saving is the interesting part. Fewer tokens per task means lower cost and faster agent runs if you host it yourself. The lead over GLM-5.3 on Terminal-Bench is only one point, though, so it is close to a tie.

Mistral also added low and medium reasoning modes alongside high. The high mode is reported as 50% more efficient than before.

Why the cybersecurity angle matters

Mistral is pitching the model for “sovereign cyber defense.” Its VP of Science, Pierre Stock, told TechCrunch that open weights mean access “can’t be cut off by definition.” In plain terms, a government or company can run it on its own servers.

That contrasts with gated releases. Google, for example, limited access to its security model, as we covered in Gemini 4 Carbon, which only cyber defenders can use for now.

The flip side is obvious. Anyone can download open weights, including attackers. Mistral says the risk of rogue agents is “solvable,” but it hasn’t shown how.

Can you run Mistral Large 4 at home?

Almost certainly not. Even the quantised build needs four Nvidia B200 or B300 data-center GPUs, which cost far more than any gaming PC. Mistral trained the model on 3,800 Nvidia Grace Blackwell GPUs in Europe.

Most people will use it through Mistral’s API once pricing appears, or through a cloud host. Mistral says smaller distilled versions are coming in the next few months, including one meant for a laptop, but it hasn’t committed to dates.

If you want a model you can run locally today, smaller options exist. Our look at Nemotron 3.5 Transcribe shows how to check whether your PC can handle a free model.

What we still don’t know

Several details are missing from the launch coverage. The reports we checked do not state the license, the API price, or the exact model size of each quantised build.

That matters because a license can limit commercial use, and a price decides whether the API is a bargain or not. Mistral has also not shared full benchmark scores for the Artificial Analysis Coding Agent Index, only that it ranks first among open models.

Independent testers will likely publish their own results in the coming days. Those numbers will tell you more than any launch chart.

Frequently Asked Questions

Is Mistral Large 4 free?

The weights are free to download from Hugging Face, but running them costs real hardware money. API pricing hasn’t been published yet.

How big is Mistral Large 4?

It has 1 trillion parameters in total, with 52 billion active for each task.

Can it run on a laptop?

Not this version. A quantised build needs four Nvidia B200 or B300 GPUs. Smaller laptop-friendly versions are planned but undated.

Is Mistral Large 4 better than GLM-5.3?

Mistral says it matches GLM-5.3 on coding while using 47% fewer tokens. Independent testing is still needed to confirm that.

If you run a company with strict data rules, Mistral Large 4 is worth testing, especially if you need a model that stays on your own servers. For everyone else, wait for the smaller versions and real pricing. Until independent benchmarks land, treat Mistral’s numbers as promising rather than proven.