Mistral Large 4, the trillion-parameter bet from Europe
Announced on October 6, 2026, this 1-trillion-parameter model is promised with open weights by month's end. What is certain and what is still a promise, separated.
Key takeaways
- Mistral Large 4 was announced on October 6, 2026: around 1 trillion parameters with a mixture-of-experts architecture, API preview already live.
- Open weights are PROMISED for the end of October 2026, not released. Distinguish announcement from release.
- Its license has not been named yet; benchmark claims so far come from Mistral itself.
- What it means for you: one more frontier-class open-weight option, with the promise of greater data control.
What happened
Mistral AI announced Mistral Large 4 on October 6, 2026. It is their flagship model: around 1 trillion total parameters with a mixture-of-experts architecture, of which only about 49 to 52 billion parameters are active per token, according to Mistral's own materials. The model is multimodal: it takes text and images in, and outputs text.
Its API preview is already live through Mistral Studio. What does not exist yet: the open weights. Mistral promises a weight release by the end of October 2026, with several reports pointing at October 27. Until the weights are actually downloadable, the status remains a promise, not a release.
What still needs proof
Two things. First, the license. Mistral's documentation tags its license merely as "Open", with no specific license named. Its predecessor, Mistral Large 3, shipped under Apache 2.0, but for Large 4 there is no confirmation. "Open-weight" and "open-source" are not the same thing; read the terms when the weights land.
Second, the capability claims. The benchmark numbers circulating are Mistral's own claims, including claimed wins on their visual-grounding, finance, and legal-agent benchmarks. Independent verification is only possible once the weights are public and third parties can test them themselves.
For context: in the same week, Reflection AI also announced Beam, a 501-billion-parameter MoE model planned under the Apache 2.0 license. October 2026 is indeed the month of open bets.
What it means for you
If the weight promise is kept, this adds one more frontier-class option you can self-host. For companies, that means greater control over data and infrastructure, without renting a closed model through an API. For developers, it means a new foundation for experiments, fine-tuning, and local products.
INTERPRETATION: the real value of this release is not its benchmark ranking, but the choice it creates. Closed models offer convenience; open weights offer control. Large 4, if the promise is kept, widens the second choice.
