Some AI labs name their models after constellations or poets. Mistral called its new flagship "Le Chonk," and at a trillion parameters, the name fits.
A Trillion Parameters You Can Download
On October 29, Mistral released open weights for Mistral Large 4, a natively multimodal mixture-of-experts model with one trillion parameters. The weights are on Hugging Face under what Mistral calls a modified MIT license: free to use, fine-tune and commercialize, unless you run a model-as-a-service business, which needs a separate commercial license. It's also available through Mistral's API, OpenRouter, Microsoft Foundry and NVIDIA-hosted endpoints, at $0.70 per million input tokens and $2.10 per million output tokens.
Mistral is targeting two jobs: software development and defensive cybersecurity. It says the model can find, reproduce, validate and patch vulnerabilities, and that it refused malicious cyber requests 96.6% of the time on average, the highest rate Mistral measured among open models. On Terminal-Bench 4.0 it scored 41%, up 13 points from the preview. Those are Mistral's own numbers, and the full technical report is "coming soon."
Sovereignty Is the Real Product
The benchmarks are good, but control is the bigger selling point. Mistral says the model was trained from scratch on European infrastructure, and open weights mean banks, governments and security teams can run it on-premises without sending sensitive code or incident data to someone else's cloud. A Mistral executive told TechCrunch the appeal is access "that can't be cut off by definition."
Mistral is also upfront about the risk. A model that's good at finding vulnerabilities can help attackers too, and the company says that while it saw the model refuse attacks on live targets, it also sees "a potential for uplift to conduct real attacks" and is adjusting its guardrails. A more aggressive pentesting variant is limited to vetted design partners.
Businesses that wanted capable AI without handing their data to a closed lab now have a serious open option. The catch is that you need the hardware and the know-how to run it yourself.
Weighing open-weight models against hosted APIs for your own AI rollout? Our free AI implementation guide covers the decisions to make before you pick a model.
Source: TechCrunch