Mistral launched Large 4 on October 6 as a multimodal flagship aimed at American closed models and Chinese open-weight rivals. Nicknamed Le Chonk, it has one trillion parameters. Builders can try a public API preview now, while the promised weight release remains ahead.
TechCrunch reports that access initially runs through a guarded endpoint. Mistral VP Science Pierre Stock told the outlet that the company plans to release weights after three weeks of safety testing with trusted partners and governments. He named cybersecurity, finance and chip design as optimized uses.
A large model with a staged release
Mistral’s announcement specifies 49 billion active parameters and targets the end of October for weights. It says training used 3,800 Nvidia Grace Blackwell GPUs in its European datacenters, which also serve the preview. TechCrunch quotes Stock’s rounded figure of 4,000 GPUs.
The deployment pitch is control. TechCrunch describes Mistral’s target audience as enterprises and institutions concerned about security and access to models. Stock argues that open weights make auditing easier. For teams choosing a provider, today’s endpoint access and the planned weight release are separate stages of that offer.
The published comparisons narrow the performance claim
TechCrunch describes benchmark results as pending. Mistral’s release page now includes reported evaluations, so readers have more evidence to assess than that report alone provides.
The lead chart shows a blind coding-quality evaluation Mistral ran with Surge AI. Large 4 scored 3.74 on a five-point scale, behind Claude Opus 5 at 4.22 and ahead of the tested Kimi and GLM models. Mistral also reports a visual-grounding score of 42% on Dense 200 against GPT-6 Astra’s 41%.
These are comparisons published by Mistral. They support a narrower claim of competitiveness on selected tasks than a general lead over closed and open rivals.
Sources
- TechCrunch: Mistral’s new 1T model aims to leapfrog closed and open rivals
- Mistral: Introducing Mistral Large 4. Lead chart credit: Mistral, coding-quality evaluation with Surge AI.
