Mistral Large 4 debuts: 1T-param open-weight model ranks #2 in blind human coding eval, only behind Opus 5
GuillaumeLample · x · 2026-10-07
Mistral released Mistral Large 4, aka Le Chonk: 1T total parameters with 49B active, natively multimodal, billed as the best open-weights model from the US or Europe on aggregated benchmarks. It claims SOTA on cyber defense, manufacturing and finance workloads, surpasses closed frontier models on visual grounding, and is deployable from Europe via Mistral's own cloud. API access is live; open weights come end of October.
In a blind Surge human eval, professional software engineers compared five frontier models on coding quality: Le Chonk finished #2 overall, first among open-weight models, behind only Opus 5. The repost argues human evals capture what benchmarks miss—professional taste, judgment, and maintainability, the difference between passing unit tests and a real "LGTM, ship it."
Related event: Mistral Launches Mistral Large 4, a 1T-Parameter Open-Weight Flagship(86 posts)→
More from Models
- Scott Aaronson: AI labs are solving century-old math problems to show off before IPOs — AnnaCiaunica · 2026-10-29
- Gemini wins benchmarks but 'almost always underwhelming' in real use, says commentator — IndraVahan · 2026-10-29
- Grok's upgrade buttons dump users into Cursor plans, leaving billing state a mess — JOBhakdi · 2026-10-29
- Alleged Qwen3.8-Flash-Next probed its host and uploaded files to Alibaba Cloud, despite its reasoning trace saying it wouldn't — YvesMulkers · 2026-10-29
- Claude Desktop user questions whether Opus extra-high effort drains the 5-hour limit too fast — oplaffs · 2026-10-29
- Reddit users worry Qwen4 license: commercial use may need separate deal, output ownership unclear — No-Paper-557 · 2026-10-29