Mistral open-weights 1T-parameter Le Chonk, beating GLM 5.3 and Kimi k3 in 64% of preference evals
b_roziere · x · 2026-10-30
Mistral's Bertrand Roziere announced open weights for Mistral Large 4 (Le Chonk):
- A 1-trillion-parameter, natively multimodal model with major gains in coding and defensive cybersecurity; RL scaling gains show no saturation.
- Claimed best OSS model in human preference evaluations, beating GLM 5.3 and Kimi k3 in 64% of comparisons.
- Consistent with Mistral's FrontierCode 2.0 benchmark, which tests real-world engineering work including whether a maintainer would merge a change.
- Ships three reasoning modes for cost-performance tradeoffs, fully open-weight and European-trained for sovereign, self-hosted deployment.
Related event: Mistral Open-Sources Le Chonk, a 1T-Parameter MoE Model(38 posts)→
More from coding & agent
- Exo: Berkeley's self-improving agent architecture cuts costs 31-62% vs Codex — arena · 2026-10-30
- Indie dev builds Backscroll: a steerable Twitter/IG feed reranked by dirt-cheap decision models — sdand · 2026-10-30
- "It was Supabase": a viral app's creator had no idea where it was hosted — gabe_ragland · 2026-10-30
- New paper asks why teams of AI agents fail where a single agent often matches them — JoshPurtell · 2026-10-30
- His app is live with users — he has no idea where it's hosted: "Codex" — gabe_ragland · 2026-10-30
- Superconductor launches: multiplayer AI workspace putting teams and coding agents in one session — sergeykarayev · 2026-10-30