France’s Mistral AI has done something its bigger rivals will not do. On October 29, the Paris lab released the full weights of Mistral Large 4, its largest model yet, so anyone can download it and run it on their own hardware. No waitlist, no API-only gate, no provider sitting between you and the model.
The release closes a promise Mistral made three weeks earlier. Large 4 first appeared on October 6 as an API-only preview, with the weights promised by the end of October. The version now available is not that preview build. Mistral says three more weeks of reinforcement learning, plus testing with governments and corporate users, improved it substantially, and Reuters reports the company describing it as its most capable openly available model for cybersecurity work.
One trillion parameters, 52 billion at work
Large 4 is a mixture-of-experts model with about 1 trillion parameters in total, of which roughly 52 billion activate for any given token. That design is the whole point: it carries the knowledge of a giant model while costing far less to run than its headline size suggests. It is natively multimodal, handling text, images and code, works in more than 160 languages, and takes up to 1 million tokens of context in a single pass, enough for very large codebases or document sets.
Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own data centers in Europe. That detail matters to the pitch. Where American labs ask customers to trust their cloud, and Chinese open models raise supply chain questions for some buyers, Mistral is selling European AI that organizations can run themselves, under their own rules.
Benchmarks that jump, with one caveat
Mistral’s own numbers show the three extra weeks of training did real work. On Terminal-Bench 4.0, a test of agentic coding, the model scores 41 percent, up from 28 percent in the October 6 preview. On FrontierCode 2.0, built by Cognition to test whether model-written code would actually be accepted by a project maintainer, Mistral claims 19.7 percent, the top score among open-weight models. On the Artificial Analysis Coding Agent Index it reports 48 percent, effectively tied with GLM-5.3 while using 47 percent fewer tokens per task.
On security, the numbers are just as aggressive. Mistral reports 51 percent on the Artificial Analysis Cyber Index and an average 96.6 percent refusal rate on malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm, the highest it has measured among open models. The caveat applies to all of these figures: they come from Mistral itself, and independent validation will take time as researchers run the downloadable weights themselves.
Half the API price of the preview
For teams that would rather not self-host, the API price has also fallen sharply. Mistral Large 4 now costs $0.70 per million input tokens and $2.10 per million output tokens, roughly half the $1.36 and $4.18 charged at preview. That undercuts the $2 and $10 per million tokens that Google charges for Gemini 4 Carbon, its own cybersecurity-focused frontier model launched this week, though Carbon is currently gated to selected defenders while Large 4 is downloadable by anyone.
| Model | Access | Input price per million tokens | Output price per million tokens |
|---|---|---|---|
| Mistral Large 4 | Open weights, plus API | $0.70 | $2.10 |
| Google Gemini 4 Carbon | Gated, selected cyber defenders first | $2.00 | $10.00 |
| OpenAI GPT-6.1 Terra | API | About half its predecessor’s price | See our GPT-6.1 Terra coverage |
The license is a modified MIT license: free to use, fine-tune and commercialize, with one carve-out. Companies that resell the model itself as a service need a separate commercial license. Distribution is wide from day one, with weights on Hugging Face and availability through Mistral Studio, OpenRouter, Nvidia NIM containers and Microsoft Foundry as the first cloud provider.
Open weights cut both ways, and Mistral knows it
Mistral’s vice president of science, Pierre Stock, put the philosophy bluntly to Reuters: “I’m not comfortable living in a world in which open weights do not exist.” Once the model is out, he added, “it’s impossible to contain access.” That protects users from being cut off by a provider, and it also means anyone can use it, including for the offensive security work the model is good at. Mistral’s answer is that defensive teams need the same tools, running under their own policies, instead of behind a provider’s refusal filters.
That is the real contrast with this week’s other big security model. Google’s Gemini 4 Carbon, covered in our launch report, is being handed first to vetted defenders through the Fairwind program while Google runs voluntary government testing. Mistral is doing the opposite: publish everything, let defenders and attackers alike have it, and bet that openness helps the defenders more.
What this means for you
If you run a security team, Large 4 is worth a serious evaluation, not because of the benchmark claims but because of what self-hosting changes. You can scan your own code with a frontier-class model without sending that code to a third party, and you can keep running it even if a provider changes its policies. The practical catch is hardware: the quantized version needs datacenter-class Nvidia GPUs, with reports pointing to as few as four B200 or B300 cards, so this is a server-room model, not a laptop one.
If you are a developer, the cheaper API and the Hugging Face weights make Large 4 one of the lowest-cost ways to test a trillion-parameter-class model right now. Just treat the benchmark table as a starting point, not a verdict. The open weights mean independent scores will arrive within weeks, and those are the numbers worth building on.
Sources: Reuters via Northland News Radio, October 29, 2026; Mistral AI open weights announcement, October 29, 2026; independent coverage of the release’s benchmarks and pricing.
