Summary
Mistral has launched a public preview of Mistral Large 4, nicknamed Lechonk, a one-trillion-parameter open-weight model trained from scratch in Europe rather than built on a Chinese open-source base. This video breaks down what the release includes and where it fits in the global model race.
The model is a mixture-of-experts system with 49 billion active parameters, natively multimodal input, and a hybrid of instruction, reasoning, and agentic capabilities. Mistral highlights strong cybersecurity performance, including results on CyberGym. API pricing is listed at $1.36 per million input tokens and $4.18 per million output tokens. It was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own European data center, and the weights are promised by the end of the month.
The host places the release in context: it is competitive among open models globally, though a Chinese open-source model still tends to beat it, and US closed-source labs remain at least six months ahead with releases like GPT 6.1 Soul and Opus 5.5. He also shares his experience trying the model in OpenCode, where runaway reasoning tokens filled the context window, and argues that open-source models will struggle to reach mainstream users until they are as easy to use as closed products.
📺 Source: Matthew Berman · Published October 07, 2026
🏷️ Format: News Analysis
