What changed
On October 6, 2026, Mistral AI opened a public API preview of Mistral Large 4, nicknamed Le Chonk, a roughly 1.05-trillion-parameter mixture-of-experts model with about 49 billion parameters active per token, native multimodal inputs and a reported million-token context window. Mistral says it will release the weights near the end of October. At launch the weights were not downloadable, so the model was not yet independently self-hostable. Vendor benchmarks and published API prices are preliminary and may change during preview.
Why it matters
Builders evaluating alternatives to closed US and Chinese models can test Large 4 through a European-hosted API before committing to infrastructure. But a promised open-weight release is different from an available licensed artifact; operators cannot yet assess self-hosting memory, deployment reliability or final licence terms. Mistral's performance claims should be checked against independent evaluations and actual workload cost.
An unusually large multimodal mixture of experts
Mistral describes a model with about 1.05T total parameters and 49B active per token, with text and image input. Those are architectural vendor claims; parameter count alone does not establish quality or efficiency.
Preview API now; open weights later
Mistral Studio serves the public preview immediately, while Mistral schedules weights for late October. Treat self-hosting, redistribution, exact licence terms and final model files as future milestones until artifacts are actually available.
Benchmarks vary by workload and test harness
Mistral claims strong coding, cyber, finance and legal results. Independent benchmark summaries show a mixed picture depending on the evaluator, prompt setup and reasoning effort. No single result establishes universal superiority; reproduce representative tasks and costs before production decisions.
Pricing and operational constraints are not final
Public provider listings showed launch or discounted rates, but may differ by provider and change during preview. Verify the actual model identifier, context and output limits, data handling, caching and invoice rates in the endpoint being used. Do not estimate self-hosting economics before weights, quantisation and hardware support are available.
What changes at weight release
If the weights and licence arrive as planned, developers can evaluate local or sovereign deployments and compare inference throughput against API pricing. The announced date is a watch item, not a completed release.