On October 29, 2026, French AI lab Mistral published the open weights for Mistral Large 4, a model with roughly 1.05 trillion parameters designed specifically for cybersecurity work. The release came three weeks after an API preview, following an additional reinforcement-learning phase that the company says significantly improved real-world performance.
A Sparse Giant Built for Security Teams
Mistral Large 4 is a sparse mixture-of-experts model. It has 1.05 trillion total parameters, but activates only around 52 billion per forward pass — keeping inference costs manageable while preserving the capacity advantage of a much larger model. The context window sits at one million tokens, long enough to ingest an entire codebase in a single request.
The security-specific benchmarks Mistral published are striking. According to the company, the model refused 96.6% of tested malicious cybersecurity requests. On CyberGym-E2E — a benchmark centered on finding and patching real memory-safety bugs in running programs — it scored 79%, ahead of DeepSeek V4.1 Flash at 76% and matching MiMo-V2.6-Pro.
Open Weights, On-Premises Deployment
The weights are now on Hugging Face under an open license, and the model is also accessible via Mistral's commercial API. Pierre Stock, Mistral's VP of Science, framed the open-weights release explicitly around model sovereignty: regulated organizations and security teams can run the full model on their own hardware, without routing sensitive data through a third-party API.
That framing is deliberate. Most labs competing in this space keep frontier model weights proprietary. Mistral is betting that the security and compliance sectors — where data-handling rules are strictest — will pay a premium for something they can audit, deploy air-gapped, and control completely.
The Reinforcement Learning Boost
Before releasing weights, Mistral ran an additional RL phase targeting terminal task execution. The result: the model's Terminal-Bench 4.0 score jumped from 28% to 41%. Terminal-Bench measures a model's ability to complete practical shell and system tasks — directly relevant to threat hunting, incident response, and automated patching workflows.
Why This Matters
The release marks a shift in who frontier-scale AI is being built for. The conventional wisdom has been that security is a use case you bolt on after building for productivity. Mistral is structuring an entire flagship release around the assumption that security teams are a first-class customer — one willing to run their own infrastructure if it means genuine control over what the model sees.
For organizations that have been waiting on the sidelines because they couldn't send proprietary data to a cloud API, a trillion-parameter open-weights model changes the calculation significantly. The benchmark numbers are self-reported, and independent verification will follow — but the model is out, the weights are downloadable, and the clock on that verification has already started.