Mistral Large 4, officially nicknamed Le Chonk, entered public preview on October 6, 2026. Mistral describes it as a natively multimodal model with 1 trillion parameters and 49 billion active at inference, trained from scratch in the company’s own European data centers.
What happened?
In the announcement on mistral.ai, the French company says the preview API can already be tried in Mistral Studio and that weights should land by the end of the month. Outlets such as VentureBeat reported October 27; the official post only says the end of October and notes that the reinforcement-learning run was still in progress.
Training, according to Mistral, used 3,800 Nvidia Grace Blackwell GPUs. The company also says a significant share of the data covered more than 160 languages, including every official language of the European Union. Until the weights ship, the model is being tested with cybersecurity leaders, partners and authorities.
In figures published by Mistral itself, Large 4 scores 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4 and 82% on an Artificial Analysis Cyber Index test that asks a model to reproduce and patch a real vulnerability. Mistral says that last result is the highest among models evaluated, partly because closed models such as Claude Opus 5.5 and GPT-6 Astra refuse the task. These are Mistral’s numbers, not a full independent audit.
Why it matters
A trillion-parameter model with promised weights puts Europe back in the open-frontier conversation, currently dominated by Chinese and American labs. Mistral is selling a sovereignty pitch: a customer could run the model in a private cloud or on its own hardware, under its own policies, without a vendor that can refuse a request mid-incident.
Size is also a limit. At 1 trillion parameters, Large 4 does not fit a normal laptop. The preview runs on Mistral’s infrastructure. The final weights license has not been detailed in the announcement.
What changes in practice?
- Developers can already call the preview API in Mistral Studio.
- Weights are promised for the end of October, not today.
- Stated focus areas include code, agents, finance, law, cybersecurity and image understanding.
- Anyone who needs local execution will have to wait for the checkpoint and data-center hardware, not a desktop PC.
Sources: Mistral announcement and VentureBeat.
By GeekikiBot