DeepSeek V4.1 benches low but feels top-tier at 4x speed, insiders say
teortaxesTex · x · 2026-09-14
- User @usrbinroygbiv reports DeepSeek V4.1 ranks near the bottom on public benchmarks, yet in personal usage he'd place it only behind astra and GLM 5.3 — while running about 4x faster, possibly due to OMP.
- teortaxesTex agrees, joking that the benchmaxxing discourse around TB 2.1 vs TB 4.0 rankings misses the point: V4.1 is almost "anti-benchmaxxed." DeepSeek clearly doesn't optimize for benchmarks or even user experience — it looks like a piece of their internal Whale AGI R&D suite.
- Note: these are unverified community observations, not official statements.
More from Models
- GPT-6 Builds a Fall Guys-Style Game in Hands-On Demo — Matthew Berman · 2026-09-14
- Playing GeoGuessr against DeepSeek 4.1 Flash: it's fairly good at geolocation — PMinervini · 2026-09-14
- Commenter argues Anthropic's book training garbles texts, 'losing' millions of books — StewartalsopIII · 2026-09-14
- InternLM releases open-source Intern-S2-397B with multimodal and agent skills, vLLM day-0 support — Nunki08 · 2026-09-14
- Best small LLMs for writing on 4GB VRAM? Reddit thread weighs the options — Mysterious-Comment94 · 2026-09-14
- Leaked Astra and Fable sizes are far below 10T, says X user: scale barely maps to capability now — teortaxesTex · 2026-09-14