Mistral Large 4 Goes Open: What a Trillion-Parameter Download Means for Your Business

On 29 October Mistral published the full weights of Mistral Large 4, its trillion-parameter flagship model, on Hugging Face. Anyone can now download the same model that powers the company's paid API and run it on their own hardware. Customers nicknamed it Le Chonk, and the name fits: this is the largest open-weight model a Western lab has ever released.

The timing matters. The strongest models from OpenAI, Anthropic and Google remain locked behind APIs, and the best downloadable models have mostly been coming from Chinese labs like DeepSeek, Zhipu and Moonshot. Mistral is making a different bet: that companies and governments want frontier-level AI they can actually hold in their hands, not just rent by the token.

For business owners, the interesting question is not who tops the benchmark table this week. It is what changes when a model of this class becomes something you can own. Let's sink in!

What Mistral Large 4 Actually Is

Mistral first showed Large 4 on 6 October as an API-only preview. Over the following three weeks the team ran additional reinforcement learning guided by customer feedback, and on 29 October it published the improved version as open weights, in FP8 and NVFP4 checkpoints, with day-one support in vLLM and other popular serving frameworks.

Under the hood it is a mixture-of-experts model with about 1.05 trillion parameters in total, of which roughly 49 billion are active for each token. In plain terms, the model stores a huge library of knowledge but only consults a small part of it per word, which keeps running costs far below what the headline number suggests. It reads text and images, writes text and code, handles a context window of one million tokens and was trained on data in more than 160 languages.

There are some practical touches too. Reasoning effort can be set to low, medium or high per request, so you pay for deep thinking only when a task needs it, and an embedded draft head speeds up generation by up to 1.9 times. The whole thing was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centers.

The Numbers Behind Le Chonk

Mistral's self-reported benchmarks put Large 4 at the top of the open-weight pack. On Terminal-Bench 4.0, which tests how well a model completes real coding tasks in a terminal, it scores 41%, just ahead of China's GLM-5.3 at 40% and well clear of DeepSeek V4.1 Flash at 31% and Kimi K3 at 21%. On the newer FrontierCode 2.0 benchmark it also edges out the same rivals.

The more commercially interesting claim is efficiency. Mistral says Large 4 matches GLM-5.3 on Artificial Analysis' Coding Agent Index while using 47% fewer tokens per task. If that holds up in practice, a model with a similar per-token price gets almost twice the work done per dollar. Pierre Stock, Mistral's VP of Science, argues that per-token rates are misleading for exactly this reason.

The usual caveat applies: nearly all of these figures come from Mistral itself, and some numbers moved between the preview and the final release. Independent testing will take a few weeks. Still, CEO Arthur Mensch is keen to point out that the company got here on roughly $6 billion raised over three years, while its competitors talk in hundreds of billions.

Why Open Weights Matter More Than Benchmarks

Benchmarks age quickly. The strategic part of this release is the word open. Once weights are published, they cannot be withdrawn. No provider can raise your price overnight, deprecate the version your product depends on, or cut off your access because of a policy change or a geopolitical argument. Stock compares AI to electricity and says no company should depend on a supplier that can switch it off.

For businesses in regulated industries the argument is even more direct. A downloaded model can run entirely inside your own infrastructure, so client data never leaves your servers. For healthcare, legal, financial and public-sector work, that single property often decides whether an AI project is allowed to exist at all. European firms also get a frontier model trained and hosted in Europe, which simplifies GDPR and data-residency conversations considerably.

And there is a quieter benefit: predictability. Hosted models change under your feet, sometimes without notice, and a prompt that worked in March can behave differently in June. A pinned, self-hosted version behaves the same way tomorrow as it does today, which starts to matter once AI output is wired into production workflows.

The Cybersecurity Bet

Mistral is positioning Large 4 squarely at security work: detecting, analysing and responding to threats across code and IT infrastructure. On CyberGym-E2E, a benchmark where models find, reproduce and patch memory-safety bugs in real C and C++ projects, it scores 79%, matching the best open rival and beating DeepSeek V4.1 Flash at 76% and GLM-5.3 Flash at 74%. Some customers already use it for automated penetration testing.

The open-weights decision is part of the same story. Security teams are often the least able to send sensitive data to an external API, because the logs and configurations they analyse are exactly what you do not want leaving the building. A strong model that runs on-premises fits that market in a way no hosted service can.

Stock also offers a more philosophical defence: attackers will get these capabilities anyway, so defenders should get them at the same time rather than waiting. You can agree with that or not, but the practical consequence for businesses is that serious automated code auditing is now available to anyone with the hardware to run it.

Read the Small Print Before You Plan Around It

Before anyone plans a migration, a few details deserve attention. The license is not plain open source. Large 4 ships under the Mistral Ecosystem Open-Weight License, a modified MIT license: internal business use, modification and building products on top are all allowed, but operating the model as a paid model-as-a-service requires a separate commercial agreement with Mistral. The previous Large 3 shipped under the more permissive Apache 2.0, so this is a slight tightening.

Hardware is the bigger filter. Even the quantised version needs four Nvidia B200 or B300 GPUs, a setup that costs well into six figures. All trillion parameters have to sit in memory even though only 49 billion work at any moment. Mistral promises smaller distilled versions, including one that runs on a laptop, in the coming months, but there are no firm dates yet.

API pricing, for those who stay hosted, is $0.68 per million input tokens and $2.09 per million output tokens, about 39% more per output token than Large 3, with the token-efficiency claim meant to offset the difference.

What This Means for Businesses Building Digital Products

First, even if you never self-host, open weights change your negotiating position. Hosted versions of Large 4 will appear on multiple providers within weeks, which pushes prices down the way it did for earlier open models. When your AI costs sit with a single closed vendor, every renewal is their conversation; when a comparable open model exists, it becomes yours.

Second, build model-agnostic from the start. If your product calls an AI model, put a thin abstraction layer between your code and the provider, so you can swap models without rewriting the application. Releases like this one are exactly why: the best-value option now changes every few months, and switching costs decide who actually benefits from that.

Third, if you sell software to clients who refuse cloud AI, such as hospitals, banks, law firms or government bodies, a self-hosted deployment option has just become a realistic line in your proposal rather than a research project. Renting GPU capacity by the hour to pilot it is a sensible middle step before anyone buys hardware.

Finally, measure cost per completed task, not per token. The 47% efficiency claim is a useful reminder that a cheaper-looking model that rambles can cost more than a pricier one that gets to the point. Run your own real workloads against two or three candidates and let the totals decide.

Final Notes

Mistral Large 4 will not replace anyone's API subscription tomorrow, and the hardware bill keeps full self-hosting out of reach for most smaller firms for now. But the release changes what is possible: a frontier-class model that a business can own, audit, pin to a version and run next to its own data. That option has value even for companies that never exercise it, because every vendor now has to price against it.

If you are weighing where AI should live in your product, behind someone's API or inside your own infrastructure, this is a good moment to look at the numbers again. And if you want help doing the maths for your particular case, that is what we do.