Imagine a building full of experts. Every time a question comes in, only a few experts are called to answer. The rest stay in the building, but do not work on that question. That is roughly how Mistral Large 4 works.
This AI model from France totals around 1 trillion parameters. But only a small fraction is active for a single token. That is why the model can be that huge on paper, yet still make sense to run on a powerful server. And now, its weights can be downloaded.
The gist, first
- Mistral Large 4 was first shown to the public on 6 October 2026.
- Its weights were released on 29 October 2026, after three additional weeks of reinforcement learning.
- Total parameters are around 1 trillion, with a 1-million-token context, and text plus image input.
- Its licence is a modified MIT called the Mistral Ecosystem Open-Weight License.
- You may use, modify, fine-tune, and sell products built from this model, except if you offer the model itself as a service, which needs a separate commercial licence from Mistral.
What does 1 trillion actually mean?
Parameters are, simply put, small knobs that determine how a model answers. The more parameters, the more a model can learn, but the heavier it is to run.
In Large 4, the total is around 1 trillion. The active-per-token figure, according to the preview announcement, is around 49 billion. But the repository name points to a figure around 52 billion. Both figures appear in reports, so this article writes it as it is: the total is 1 trillion, the active figure sits in that range, depending on which source you read. One uncertain figure must not be written as if it were settled.
This model also uses a mixture-of-experts architecture. Back to the building analogy: there are 128 routed experts, and only four experts plus one shared expert are active per token, according to circulating technical reports. So this model does not fire up all of its parts every time it answers.
A 1-million-token context is also huge. Tokens are chunks of words and symbols. With that much context, the model can read a large codebase or a long document in one session, without having to keep cutting it up. For developers and security teams, that is what makes this model interesting, not just its big name.
Open-weight is not the same as open source
This is the most commonly misunderstood part. Open-weight means the model’s weights can be downloaded and run yourself. It does not mean everything about the model is free of rules.
Mistral’s licence here allows use, modification, fine-tuning, and commercialisation. But there is one condition to remember: if your business is offering this model as a service, you must first obtain a separate commercial licence from Mistral. So using it inside your own product is a different story from selling access to the model itself.
For companies that want to self-host, this is still good news. You can run it on your own server, in the region you choose, including in Europe, without always going through someone else’s API. But read the licence to the end before using it for business, because that single clause can change your monetisation plans.
What is this model pitched for?
Mistral positions Large 4 for coding and cybersecurity. On the coding side, Mistral claims a score of 48 percent on the Artificial Analysis Coding Agent Index, claimed to be on par with GLM-5.3 as the best open model on that index, but with 47 percent fewer tokens per task. On Terminal-Bench 4.0, its score rose from 28.3 percent at preview to 41 percent after three additional weeks of reinforcement learning.
On the security side, Mistral claims 51 percent on the Artificial Analysis Cyber Index, and a 96.6 percent refusal rate for malicious cybersecurity requests. There are also claims of 82 percent on a vulnerability reproduce-and-patch test, and 93 percent on Cybench, which contains 40 challenges from security competitions.
All of these figures are Mistral’s and its partners’ numbers, not final independent test results. The full technical report was also said to be still coming soon when the weights were released. So treat these figures as an early picture, not a final verdict on who is best.
One more thing worth appreciating: the more aggressive pen-testing version was not opened freely. That version remains restricted to verified partners. In other words, Mistral is trying to open a model that is useful for defenders, without immediately releasing the version easiest to misuse.
Price and how to access it
Through Mistral’s own API, the price is 0.70 US dollars per million input tokens, 0.07 US dollars per million cached input tokens, and 2.10 US dollars per million output tokens. This price is down roughly half from the preview price, which was 1.36 US dollars for input and 4.18 US dollars for output.
The weights are on Hugging Face under the repository name mistralai/Mistral-Large-4-1T-A52B, and can be run with vLLM or SGLang. The model is also available through the Mistral API, Mistral Studio, OpenRouter, NVIDIA endpoints, and Microsoft Foundry, said to be the first cloud to carry it.
For those wanting to run it themselves, do not imagine this running on an ordinary office laptop. A 1-trillion-parameter model is heavy. By rough calculation, at 16-bit precision the weights alone are around 2 terabytes, and at 8-bit around 1 terabyte, not counting memory for the model’s working state. So self-hosting here means a server actually designed for it, not a weak phone or a thin laptop.
Who this model suits, and who it does not
It suits developer teams wanting an open-weight coding model they can self-host, security teams wanting to experiment with a model to help analyse vulnerabilities, and companies that care about where their data is processed. For them, the combination of downloadable weights, long context, and a lower API price is genuinely attractive.
It is less suited to those who just want a chatbot for everyday questions. For that need, a smaller, cheaper model via a chat app is usually enough, and you do not need to worry about servers, MaaS licences, or fine-tuning.
It is also less suited to those hoping every benchmark figure can be trusted at face value. Wait for the full technical report, and compare with independent tests before using it for big decisions.
So, is this big news or just a numbers gimmick?
This is big news for the open-weight world, especially from Europe. A 1-trillion-parameter model whose weights are opened, with a relatively permissive licence that still has a condition for MaaS, is not something that happens often. But the size of the number is no guarantee this model is the best fit for everyone.
What makes Large 4 worth watching is not just its 1 trillion figure. What is interesting is the package: long context, image input, a lower API price, and a focus on coding and security, plus the option to run it yourself. If you are on a team that actually needs that package, it is worth trying. If not, just be aware of it for now, while waiting for the technical report and independent tests to come out.
Quick Q&A
Is Mistral Large 4 free? Its weights can be downloaded and used under the licence, but a separate commercial licence applies if you offer the model as a service. Its API is paid per token.
Is open-weight the same as open source? No. Open-weight means the weights can be downloaded and run, but the licence and its rules still apply.
How many active parameters? Around 1 trillion in total. The circulating active figure is in the range of 49 billion from the preview announcement and around 52 billion from the repository name.
Are its benchmark figures final? Not yet. Those figures are from Mistral and its partners, and the full technical report is still awaited.
Sources
- Skimwire, Mistral Large 4 open weights: specs, license, cyber claims
- Pivot News, Mistral opens the weights of its 1-trillion-parameter model
- Dev.to, Mistral Releases Open Weights for Its 1-Trillion-Parameter Large 4
- Deyron Labs, Mistral releases Mistral Large 4 as open weights
The main illustration in this article is a generated illustration, not Mistral’s official logo or photo. This article was written on 1 November 2026 based on the 29 October 2026 weights release, and the benchmark figures quoted are Mistral’s and its partners’ figures, not final independent test results.
