Mistral Releases “Le Chonk” (ML4) – Open Trillion‑Parameter Model Tops Coding Benchmarks
What Happened — On 6 October 2026, French AI lab Mistral released the weights for its new large‑scale model, Mistral Large 4 (ML4), nicknamed “Le Chonk.” The 1‑trillion‑parameter model is offered in preview and its weights are publicly downloadable. Independent testing shows ML4 leading the open‑model coding agent index, improving its Terminal‑Bench 4.0 score from 28 % to 41 % and matching proprietary rivals while using 47 % fewer tokens per task. Mistral also advertises built‑in cybersecurity controls and data‑sovereignty options, positioning the model as a more defensible alternative to closed‑source offerings.
Why It Matters for Trust & Control Assurance — The release tests an organization’s AI‑governance controls: you must be able to map model‑risk objectives to your audit framework and retain defensible evidence of evaluation.
Who Is Affected — AI model developers, enterprise software vendors, and any organization that integrates large language models into products or services.
Recommended Actions —
- Map AI‑model governance objectives (risk assessment, testing, data‑handling) to your framework of record (e.g., NIST AI RMF, ISO 42001).
- Capture and retain evidence of model evaluation, token‑usage metrics, and security‑control configuration for audit readiness.
Technical Notes —
- Model size: ~1 trillion parameters; released as raw weights (no compiled binaries).
- Benchmark: Terminal‑Bench 4.0 (coding), Artificial Analysis Coding Agent Index.
- Security claims: customizable control settings, data‑sovereignty options, “expanded cybersecurity capabilities” for early partners.