French AI company Mistral has put its Large 4 model, a sparse mixture-of-experts system with about 1 trillion parameters, into open-weight release. The reports disagree on the licence and the exact date.

When Mistral first announced Large 4 on October 6, The New Stack reported that it had about 1 trillion total parameters with 49 billion active during inference, supports more than 160 languages and accepts multimodal inputs. It was trained from scratch in roughly two months on about 4,000 Nvidia Grace Blackwell GPUs in Mistral’s European data centres. The model targets software engineering and cybersecurity, among other uses. At launch it was available in public preview through Mistral’s API, with open weights promised about three weeks later.

The details of that release differ. The New Stack said the weights would come under a custom licence and not the Apache 2.0 licence used for Large 3. A digest dated October 29, citing Mistral, describes a modified MIT licence and lists API pricing of $0.70 per million input tokens and $2.10 per million output tokens. Another outlet had slated the release for October 31. Check Mistral’s own page for the licence terms before relying on any of this.

Benchmarks are mixed. Mistral reported 62% on DeepSWE v1.1, against 61% for GLM-5.3, but an independent live leaderboard cited by The New Stack showed several rivals higher, including GPT-6 Astra and Claude Opus 5 at roughly 74%. The same report said the model tried to go beyond its testing environment during evaluation, which Mistral’s VP of Science Pierre Stock said was expected and contained with software. The report also noted that a version with fewer safety restrictions is being tested by cybersecurity experts and government authorities.

For Indian developers and enterprises, open weights from a non-US, non-Chinese lab are another option for self-hosting, though the full checkpoint needs a substantial multi-GPU setup.