Mistral AI releases weights for trillion-parameter Mistral Large 4

Security teams and regulated enterprises can now download a frontier-scale model and keep sensitive code, logs and incident data inside their own infrastructure. That gives them an alternative to hosted models from OpenAI and Anthropic, whose access a provider can change. Independent benchmark results are still unknown. So are long-context accuracy and the real cost of running a trillion-parameter model.

Mistral AI opens the weights

Mistral AI put the weights on Hugging Face and serves the same model through its own API.

Mistral AI calls it Le Chonk. The model is natively multimodal and takes text and images as input. Mistral Large 4 has about 1 trillion total parameters. Only a slice runs on any given request: roughly 52 billion active parameters in a mixture-of-experts design, according to two reports. A third report puts the active count at about 49 billion and the total at about 1.05 trillion. That sparse design is what makes a model this size servable at all, because the full network does not fire for every token.

Why security teams care

Mistral AI pitched the release at security work. That work includes finding vulnerabilities, reproducing them in a sandbox and patching the code without breaking features.

Pierre Stock, Mistral AI's vice president of science, argued that downloadable weights protect against a vendor switching off access. "Once you open source the weights, they're all over the internet. It's impossible to cut off access," he told reporters. He added that "you don't want to live in a world in which your neighbor can cut off electricity from one day to another". The company reports a 51% score on the Artificial Analysis Cyber Index and a 96.6% refusal rate on tested malicious cybersecurity requests.

Dual use is the obvious objection. Rafe Pilling, senior director of threat intelligence at Sophos, called it "clearly a dual-use technology". Anything downloadable and unfettered, he said, can be abused as well as used for productive purposes.

The benchmark claims

Most of the performance numbers come from Mistral AI itself.

The company reports 48% on the Artificial Analysis Coding Agent Index. It says the model uses 47% fewer tokens per task on average than GLM-5.3. TechRepublic reported that the 48% tied GLM-5.3; lavx.hu put GLM-5.3 one point ahead, at 49. On Terminal-Bench 4.0, Mistral AI reports 41%, up from 28% at preview, a gain of 13 percentage points. It also reports 73.3% on Finch finance workflows and 84.5% on BrowseComp.

Attack resistance on the B3 Agent Security Benchmark rose to 97.1% from 93.3% in preview. The company reported the figures.

The cost of running it

The quantized checkpoint alone is 606GB. Full-precision hosting therefore needs substantial GPU capacity. The license is a modified MIT license. It permits free use, modification, fine-tuning and commercialisation. Businesses offering the model as a service need a separate commercial licence. API pricing, for teams that would rather rent, is $0.70 per million input tokens, $0.07 for cached input and $2.10 for output.

Mistral AI trained the model on Nvidia GB200 GPUs. It used its own European data centres. One account puts it at 3,840 GPUs for 45 days, about 4.15 million GPU hours; lavx.hu reports roughly 4,000 GPUs over two months. Organisations that host it still need technical expertise. They also need security processes to run and monitor it themselves. Mistral AI has also not published a long-context accuracy curve for the 1 million token window, shattered.io noted.

The open-weight contest

The nearer fight is with Chinese open-weight models.

CEO Arthur Mensch described AI as "a tool for empowerment, not dominance".

The company says enterprises and governments get "frontier AI cyber defence capabilities they can fully own and control". What remains open is whether the benchmark lead survives outside testing. It is also unclear what a full-precision deployment costs in hardware.