Mistral, the French AI company, released the open weights of Mistral Large 4 on 29 October 2026. Its flagship 1-trillion-parameter model, previewed earlier in October, can now be downloaded from Hugging Face and run on a business's own servers, or used through Mistral's API, Microsoft Foundry and NVIDIA. For most small and medium businesses the practical questions are what the licence allows, what it costs and whether anyone in their supply chain will run it themselves.
What happened
In its announcement "Mistral Large 4 (Le Chonk) is now open weight", Mistral says the model that went into public preview earlier in October is now open-weight. Open-weight means the trained model files are published, so anyone with suitable hardware can download them and run the model without sending data to Mistral.
TechCrunch reported on 29 October 2026 that the release lets businesses and governments download and deploy the system themselves, with the weights on Hugging Face and through Mistral's API. We covered the preview of Mistral Large 4 when it first appeared; this is the step where the files themselves became public.
Mistral describes the model as a hybrid instruct-and-reasoning model: one model that can answer quickly or work through a problem step by step. It takes text and images as input and writes text. Mistral says it was trained from scratch on European infrastructure and is natively fluent in 160+ languages, which is part of how the company pitches it to governments and regulated industries that want to know where their AI comes from.
The release also comes with a caution. Mistral says that since the preview it watched cybersecurity-related use in its live traffic, that it sees a potential for uplift to conduct real attacks, and that it is modifying its guardrails accordingly. A specialised cyber variant for penetration testing and red-teaming stays in preview for vetted enterprise design partners, under contractual safeguards.
Key details
| Item | What Mistral and the model card say |
|---|---|
| Release of the weights | 29 October 2026, after a public preview earlier in October |
| Size | 1 trillion parameters in total; 49 billion active per token (52 billion including embeddings and output layers) |
| Design | Mixture-of-Experts: 128 routed experts, 4 active per token, plus 1 shared expert |
| Context length | 1M tokens |
| Input and output | Text and images in, text out; three reasoning efforts: low, medium and high (the default) |
| API price | $0.70 per million input tokens, $0.07 cached input, $2.10 output |
| Where to use it | Mistral Studio, the Mistral API, OpenRouter, NVIDIA-hosted endpoints and NIM containers, Microsoft Foundry, or your own servers |
| Licence | Mistral Ecosystem Open-Weight license, a modified MIT license; a Model-as-a-Service business needs a separate commercial license |
The Hugging Face model card adds the engineering detail. Mixture-of-Experts means the model is split into many specialist parts and only a few of them work on each word, which is why only 49 billion of its 1 trillion parameters are active at a time. The card says the model was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe.
On performance, Mistral reports 48% on the Artificial Analysis Coding Agent Index, which it calls effectively tied with GLM-5.3, an open-weight model TechCrunch describes as a Chinese competitor, while using 47% fewer tokens per task on average. It also reports a 13 percentage point improvement to 41% on Terminal-Bench 4.0 compared with the preview, and 51% on the Artificial Analysis Cyber Index, a test of finding, reproducing and patching software vulnerabilities. These are Mistral's own published results.
Why it matters
Many of the most capable open-weight models of the past year came from Chinese labs, such as the Kimi K3 open weights and DeepSeek's releases. Mistral Large 4 gives organisations a large open-weight option trained in Europe, which matters to buyers who have to answer questions about where a model came from and who can switch it off.
Fewer tokens per task also matters in money terms. Most AI services charge by the token, so a model that reaches the same answer with less text costs less per job, whether you pay Mistral's API or pay for your own servers' running time.
The security angle cuts both ways. A model that is good at finding and patching vulnerabilities helps defenders, and Mistral says its guardrails refused malicious cyber requests at an average rate of 96.6% and resisted 97.1% of attacks on the B3 Agent Security Benchmark, up from 93.3% in the preview. But once weights are public, anyone can run them, which is why Mistral says it keeps adjusting those guardrails. It is a reminder that attackers use AI too: the ASD's advice on protecting AI accounts and API keys applies to every business that pays for an AI service.
What this means for businesses
Very few small businesses will run Mistral Large 4 on their own hardware. The model card's serving examples use a node of 8 GPUs, and the card gives no exact memory figure, so this is a job for a data centre or a cloud provider, not an office server. A smaller quantized checkpoint exists, but it still targets NVIDIA's Blackwell data-centre GPUs.
Where it can matter to you is through suppliers and through the API:
- Ask your software vendors which model sits behind their AI features. Some may move to Mistral Large 4 because they can host it themselves. Ask where it runs, which country your data is processed in, and whether your data is used for training. The same questions apply to AI features in hiring tools, from applicant screening to candidate verification platforms such as Talent Tabloid.
- Check the licence if you sell AI to others. If your business offers a hosted AI service built on this model, Mistral says you need a separate commercial license before you use it commercially.
- Compare on your own work. Run the same real tasks, such as summarising a contract or drafting a reply to a quote request, through your current tool and through Mistral's API, and compare quality and cost per task.
- Keep the security basics in place. An open-weight model on someone's server is only as safe as that server. Ask for access controls, logging and patching in writing, as you would for any hosted system.
- Treat AI-written code with care. If a developer uses this model for coding, the checks in our guide to whether AI-generated code is secure still apply before anything goes live.
If you are weighing up which AI model or provider fits your systems, our IT consulting work includes vendor and software selection, and our cyber security team can review how an AI service connects to your data. Where a self-hosted model makes sense, cloud hosting is usually the realistic route. Ask us for a free first consultation and a written quote before you commit to a platform.
Frequently asked questions
Can a small business download and use Mistral Large 4 for free?
The weights can be downloaded from Hugging Face and used commercially under the Mistral Ecosystem Open-Weight license, a modified MIT license. The catch is hardware: the model card's serving examples use a node of 8 GPUs, so most small businesses will reach it through Mistral's paid API or a cloud platform instead.
How much does Mistral Large 4 cost on the API?
Mistral lists $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. A token is a small piece of a word, so the bill depends on how much text your software sends and how long the answers are.
Does running it ourselves keep our data private?
If the model runs on servers you or your provider control, your prompts and documents do not go to Mistral. You still own the security of those servers, the access controls and the logs, so the privacy benefit depends on how well that environment is run.
Is the open-weight licence the same as open source?
Not quite. Mistral calls it open-weight and releases it under a modified MIT license that allows commercial and non-commercial use, but a business offering the model to others as a hosted service must get a separate commercial license from Mistral first.
Should we switch from the AI tools we already use?
Not on benchmark scores alone. Mistral's figures are its own measurements, so test the model on your real documents and tasks next to your current tool, and compare quality, cost and where the data goes before changing anything.
Key takeaways
- Mistral released the open weights of Mistral Large 4, a 1-trillion-parameter model, on 29 October 2026.
- The licence allows commercial use, but a Model-as-a-Service business needs a separate commercial license.
- Self-hosting needs data-centre hardware; most small businesses will use the API or a cloud platform.
- API pricing is $0.70 per million input tokens and $2.10 per million output tokens.
- Test it on your own tasks and ask suppliers where any AI model processes your data.