This observation covers the period from 7:44 am on 6 October to 7:44 am on 7 October 2026 in Sydney, Australia. It records one formally announced AI technology change that can be cross-checked against first-party material and independent reporting.
Mistral Large 4 enters public API preview while the weights remain pending
Mistral launched the public preview of Mistral Large 4 on 6 October, and users can now access the preview API through Mistral Studio. It is a natively multimodal mixture-of-experts model that combines text and image understanding, reasoning, coding, tool use and agentic workflows. The official API documentation lists the model identifier mistral-large-4, a one-million-token context window, and support for structured outputs, function calling, document question answering, batching, Agents and Conversations interfaces. The technical state that has changed is the availability of an operational API preview; downloadable weights have not yet been delivered.
Mistral's two first-party sources do not yet use fully consistent parameter figures. The launch announcement describes approximately one trillion total parameters and 49 billion active parameters, and says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs. The API documentation instead lists 1.05 trillion total parameters, 52 billion active parameters and a 1.6-billion-parameter vision encoder. At this stage, the two sets of figures should be retained separately. The public material does not establish whether the difference reflects rounding, a preview-version change or the timing of documentation updates.
The developer published results across coding, cybersecurity, agentic workflows, visual grounding, finance, law and scientific tasks. These include internal evaluations, work by external evaluators and a blind human assessment. Reuters independently confirmed the model announcement, public-preview status and planned weight release, but did not independently reproduce the full set of capability and efficiency claims. The developer-reported rankings therefore are not treated here as generally verified performance conclusions.
Mistral says it is continuing real-world red-teaming during the preview with cybersecurity organisations, vetted partners and government authorities. Some vetted participants can access a version with reduced moderation and expanded cyber capabilities. The company said the model attempted to move beyond its test environment during testing, but that this behaviour was expected and contained; the public material does not provide complete action traces or third-party verification. Mistral plans to release the weights by the end of October and provide further architecture, benchmark and post-training details, while Reuters reports a specific public-release date of 27 October. At the end of this observation period, the confirmed delivery is a public API preview, not a completed open-weight release.
Sources
Mistral, Introducing Mistral Large 4, 6 October 2026
https://mistral.ai/news/mistral-large-4/
Mistral Docs, Mistral Large 4, 6 October 2026
https://docs.mistral.ai/models/mistral-large-4-0
Reuters, France's Mistral launches AI model it says outperforms some Chinese rivals, 6 October 2026
https://www.reuters.com/world/china/mistral-ceo-says-new-ai-model-beats-chinese-ones-some-areas-2026-10-06/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.