Mistral Releases Open Weights for 1-Trillion Parameter Mistral Large 4
Mistral AI has published the weights for Mistral Large 4, delivering frontier coding results, defensive cybersecurity guardrails, and three flexible reasoning tiers.
Mistral AI has officially released the weights for Mistral Large 4, a 1-trillion-parameter multimodal model codenamed "Le Chonk." Following an initial public preview earlier this month and three weeks of further reinforcement learning, the European AI lab has made the model's weights openly accessible, emphasizing advancements in automated programming, defensive cybersecurity, and infrastructure sovereignty.
Built and trained entirely within Europe, the model is targeted at enterprises and developers seeking self-hosted or cloud-based deployments without relying strictly on black-box external APIs. Alongside the weights release, Mistral introduced three distinct reasoning settings—Low, Medium, and High—allowing teams to balance inference speed, cost, and analytical depth depending on task complexity.
High coding accuracy paired with token efficiency
A primary focus of the post-preview reinforcement learning was agentic programming. According to benchmark figures released by Mistral, Mistral Large 4 achieved an open-weight leading score of 19.7% on Cognition’s private FrontierCode 2.0 evaluation, which tests whether real repository maintainers would accept and merge an agent's pull requests.
On the Artificial Analysis Coding Agent Index, the model posted a 48% score, tying rival open-weight model GLM-5.3 while using 47% fewer tokens per task on average. In command-line environments evaluated under Terminal-Bench 4.0, the model reached 41%, marking a 13-percentage-point increase over its preview iteration alongside a 30% reduction in generated tokens.
Blind human evaluations run by Surge Code showed evaluators preferring Mistral Large 4 over other open-weight models including GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, placing it behind only Claude Opus 5.5 among the systems tested.
Defensive security and enterprise capabilities
Beyond development tools, Mistral positioned the release toward cybersecurity defense. On the Artificial Analysis Cyber Index—evaluating vulnerability identification, reproduction, and remediation—Mistral Large 4 scored 51%. The team noted that the model processes these legitimate vulnerability-handling workflows without triggering false-positive safety refusals, while maintaining a 96.6% refusal rate against explicitly malicious cyber requests across benchmarks such as JailbreakBench, StrongREJECT, and AgentHarm.
Mistral confirmed it is collaborating with design partners like Orange Cyberdefense on controlled red-teaming and penetration-testing workflows. On the B3 Agent Security Benchmark, the system resisted 97.1% of adversarial prompt attacks, up from 93.3% during the preview.
Other enterprise metrics also saw gains: the model achieved 73.3% on the FinWorkBench (Finch) spreadsheet and accounting benchmark, trailing only Kimi K3 among open offerings, and reached 84.5% on the BrowseComp web navigation test.
"Technological sovereignty is not optional; it is a key prerequisite for maintaining our customers' trust," stated Zakaria Moursli, Executive Vice President at French partner La Banque Postale, highlighting the role of locally audited models for enterprise continuity.
Deployment, licensing, and pricing
Mistral Large 4 is distributed under the Mistral Ecosystem Open-Weight license, a modified MIT agreement. It permits free inspection, modification, fine-tuning, and internal or commercial usage, with one explicit exception: companies offering managed Model-as-a-Service platforms must secure a separate commercial license.
Organizations can self-host the weights using inference engines such as vLLM or SGLang, utilize NVIDIA NIM containers, or access managed deployments via Microsoft Foundry. Microsoft's Corporate Vice President Tina Schuchman noted that the integration provides enterprise customers greater choice for private cloud deployments.
For API users, Mistral set official inference pricing at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens.
What it means for developers
For software engineers building agentic workflows, the most notable aspect of Mistral Large 4 is its token efficiency. Running autonomous multi-turn loops where models plan, execute commands, inspect compiler errors, and patch code often leads to high inference bills. By cutting token consumption on benchmarks like the Coding Agent Index by nearly half compared to open competitors of similar accuracy, the model significantly reduces operational overhead.
The addition of reasoning tiers (Low, Medium, and High) gives engineers direct programmatic control over compute budgets. Simple script adjustments can run under the Low mode, while complex refactoring or security reviews can rely on High reasoning.
For developers evaluating different options across the ecosystem, developers can try top AI models cheaply through one API at https://apixoai.online.
Finally, teams subject to strict compliance or data residency rules gain an open-weight alternative that can be fine-tuned via tools like Mistral Forge and deployed entirely within private VPCs or on-premises servers, removing reliance on proprietary API vendors.
Source: Le Chonk: Built for Coders, Trusted by Defenders. — Mistral AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
