The French startup's trillion-parameter model can be downloaded and run privately, a pitch aimed at organisations wary of relying on closed AI services
French AI startup Mistral released its Large 4 model to the public on October 29. The company says it is designed to help organisations find and fix cybersecurity vulnerabilities.
The bigger story is how you get it. Closed models such as Google’s Gemini are mainly accessed through online services. Mistral Large 4 can be downloaded and run independently by anyone, which makes it unusual at this scale.
What Mistral actually shipped
Mistral Large 4, shortened to ML4 and nicknamed “Le Chonk,” has roughly 1 trillion total parameters.
Only 52 billion parameters are active for any given token, so most of the model sits idle on each step.
The model can run on as few as four Nvidia B200 or B300 GPUs. For a trillion-parameter system, that is a surprisingly modest footprint.
ML4 supports a 1M-token context window and fluency in over 160 languages.
A context window is how much text a model can consider at once. At a million tokens, it can take in very large codebases or document sets in a single pass, which matters when hunting for flaws.
The weights went live on Hugging Face under a modified MIT-style license. Organisations can deploy the model on their own infrastructure rather than routing sensitive data through someone else’s servers.
For teams that prefer not to host it, API pricing after release is approximately $0.70 per million input tokens and $2.10 per million output tokens.
The benchmark case for defenders
Mistral is pitching ML4 squarely at security work. That includes identifying and patching vulnerabilities, plus agentic coding and broader enterprise tasks.
On cybersecurity benchmarks like CyberGym-E2E, the model scored 79–82%. Its results also topped the best open-weight scores on Cybench, another security test suite.
The model improved noticeably between its October 6 preview and the full release. On Terminal-Bench 4.0, its score climbed from 28% to 41%.
Mistral attributes that jump to reinforcement learning shaped by feedback from government and corporate users.
ML4 also matches or beats some leading Chinese models, including GLM-5.3, on cybersecurity tasks while using significantly fewer tokens.
Mistral says ML4 has high refusal rates for malicious prompts. The goal is a model that is useful in adversarial settings without becoming an attacker’s assistant.
Why Mistral is betting on open weights
Mistral trained ML4 on 3,800 Nvidia Grace Blackwell GPUs located in European data centers.
The French startup has leaned hard on sovereignty. Its argument is that organisations should not depend on a third-party provider that could revoke access, change terms, or go down at the worst possible moment.
Pierre Stock, Mistral’s VP of Science, highlighted the importance of open weights in AI development. In practice, that means the customer holds the model, not just a subscription to it.
What this means for enterprises and the AI race
The most direct beneficiaries are organisations with sensitive security workloads and strict data-handling rules. ML4 offers frontier-scale capability without requiring them to ship internal code or logs off-premises.
The four-GPU minimum lowers the barrier further. Running a trillion-parameter model in-house still takes serious hardware, but it no longer requires a hyperscaler’s worth of it.
The comparison with Chinese open models like GLM-5.3 also matters. Enterprises that want open weights but have concerns about a model’s origin now have a European alternative with competitive security benchmarks.
Pricing will shape adoption among smaller teams that cannot host the model themselves. At approximately $0.70 and $2.10 per million input and output tokens, the API route gives them a way in without buying GPUs.
