Mistral Research Head on Le Chonk: Trained on 3,800 GPUs, 10-Trillion-Parameter Models Coming
mattturck · x · 2026-10-23
Matt Turck interviews Pierre Stock, who runs research at Mistral AI, on the Western open-weight comeback.
Key points:
- Le Chonk is training on 3,800 GPUs in Mistral's own datacenters and is still training
- MoE explained, roadmap toward 10-trillion-parameter models
- What Mistral borrowed from DeepSeek; one model with multiple reasoning modes; native multimodality (satellites, manuals, computer use)
- RL training on coding, cyber, finance, chips; whether RL teaches anything new
- Anecdote: a sandboxed model pinged its trainer
- Mistral Large 5 coming; comparison vs Beam, Muse Spark Open and Chinese models
Related event: Mistral Research Lead on Le Chonk and Open-Weight Revival(2 posts)→
More from Companies & People
- Paris 1 Panthéon-Sorbonne opens 2-year AI Philosophy postdoc, salary from €2,450/month — Dr_Atoosa · 2026-10-23
- Pichai: no company could have gotten its 2-year-old AI compute bets right — rohanpaul_ai · 2026-10-23
- Percy Liang teaches new MIT course on self-evolving AI agents across four evolution levels — pliang279 · 2026-10-23
- xAI launches Grok for Startups with shared credits and 6-month dollar-for-dollar match — XFreeze · 2026-10-23
- ChatGPT makes ~$130M/month on the App Store, outspending Claude and Perplexity 6x on Meta ads — fekdaoui · 2026-10-23
- NVIDIA Opens 2027-2028 Graduate Fellowship Applications: Up to $60,000 Per Student — ArashVahdat · 2026-10-23