Mistral Releases Mistral Large 4 Weights, an Open 1-Trillion-Parameter Model for Coding and Cybersecurity

What happened

Mistral released the open weights of Mistral Large 4, a 1-trillion-parameter natively multimodal model. The model entered public preview earlier this month and has now been released with improvements driven by three weeks of additional reinforcement learning focused on coding and defensive cybersecurity capabilities. The weights are now available to developers and organizations for self-hosted deployment.

Coding performance and efficiency

Mistral Large 4 leads open-weight models on multiple coding benchmarks. On Cognition's FrontierCode 2.0 benchmark, which evaluates end-to-end code quality and whether maintainers would merge an agent's changes, the model scores 19.7%. On Terminal-Bench 4.0, testing terminal-based coding workflows, it improved 13 percentage points from public preview to 41%, leading open-weight models while reducing output tokens by 30%.

Token efficiency is a key differentiator. On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48%, matching the leading open-weight model GLM-5.3, but uses 47% fewer tokens per task on average. This efficiency matters for developers running agentic coding systems at scale, where token costs accumulate quickly. In blind pairwise human evaluations by Surge Code, Mistral Large 4 was preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models tested.

To help developers balance performance and cost, Mistral Large 4 supports three reasoning modes: Low for simple tasks, Medium for everyday tasks, and High (the default) for complex use cases requiring high precision and thorough reasoning.

Cybersecurity capabilities

Mistral Large 4 scores 51% on the Artificial Analysis Cyber Index, which measures a model's ability to discover vulnerabilities in codebases, reproduce and validate them, and patch them without breaking existing functionality. The model performs these defensive security tasks without unnecessary refusals, a capability validated through collaboration with leading cybersecurity vendors who tested it in real-world environments.

Safety improvements keep pace with capability gains. The model distinguishes between defensive and malicious cybersecurity tasks, maintaining strong performance on defensive benchmarks while achieving a 96.6% refusal rate on malicious cyber requests, the highest refusal rate Mistral reports among open models. This selective refusal approach allows security teams to use the model for legitimate defensive work without blocking genuine security research and remediation.

Sovereign AI and deployment control

The model was trained from scratch on European infrastructure with clear lineage from inception to release, addressing organizational concerns about data sovereignty and model provenance. Open weights enable organizations to deploy the model on their own infrastructure, whether on-premises, in a VPC, or air-gapped environments, without routing sensitive data through external APIs.

Organizations can adapt and customize the model to their own security workflows, detection logic, and analyst feedback. Because weights are open, teams can upgrade or replace models without losing evaluations and knowledge built around previous versions. Security teams retain control over what agents can access, which tools they can use, and when human approval is required for actions.

Industry adoption and validation

Major cybersecurity vendors and enterprises have begun testing and deploying Mistral Large 4. Booz Allen demonstrated that the model matches vulnerability detection recall rates against other frontier cyber models with a lower cost curve, and supports data-sovereign deployment for clients with strict data-residency requirements. Cisco, Cloudflare, CrowdStrike, Elastic, Orange Cyberdefense, and others highlighted the value of open weights for security operations, emphasizing control, auditability, and the ability to deploy under their own infrastructure without external dependencies. Crédit Agricole is using the model to strengthen strategic autonomy with European open-weight models on its own infrastructure.

Source

AI-assisted, verified against the source. This post was drafted with AI from the primary source linked above and checked automatically against it. For anything you depend on, read the original.