Report: Mistral and others increasingly at risk of open model ‘abliteration’

Mistral AI has unveiled its latest flagship model, Mistral Large 4, available for public preview. This trillion-parameter, multimodal mixture-of-experts system is designed to tackle complex tasks such as cybersecurity and agentic reasoning. Unlike many closed foundation models, Mistral is committed to releasing the model weights to the public by the end of October 2026, allowing for self-hosting, fine-tuning, and auditing.

The model, nicknamed "Le Chonk," operates on a mixture-of-experts architecture with a reported 49 to 52 billion active parameters per token, balancing model capacity and inference costs. Built to process both text and images, Mistral Large 4 is targeted at domains like cybersecurity, finance, and agentic AI. While the preview API is now available through Mistral Studio, the model's performance is still improving during the ongoing reinforcement learning process.

However, the substantial hardware requirements for self-hosting mean that only teams with significant computational resources can currently run the model locally.