DeepSeek cut KV cache per token from 389KB to 890B in two years as open-closed gap debate rages
teortaxesTex · x · 2026-10-27
Responding to claims that the open vs closed source capability gap is widening, the author argues revenue per dollar of compute is a better frontier indicator than benchmark-month gaps: DeepSeek's gains from V3.2 to V4 to V4.1 were even steeper, though top US models still command far more absolute compute.
Charts from DeepSeek's paper show Global KV Cache Per Token falling from 389,120 bytes (DeepSeek-V1) to 48,068 (V3.2, 8.1×), 3,514 (V4-Flash, 13.7×) and 890 bytes (V4.1-Flash, 3.9×), with V4.1-Flash keeping single-token decode FLOPs nearly constant from 4K to 1024K context, slashing long-context cost.
More from Models
- OpenAI staffer explains GPT-6.1 Terra: midrange bet on vision and technical communication — goodside · 2026-10-27
- Harmonic Teases: AI Taught to Prove Math, 'Something Harder' Coming Soon — satnam6502 · 2026-10-27
- OpenAI quietly swaps Plus/Pro default to cheaper GPT-6.1 Terra and cuts Codex allowance — kimmonismus · 2026-10-27
- 'Never been this uninterested after an OpenAI model release' — theteknosaur · 2026-10-27
- Dev questions reports of Chinese models writing insecure code, citing thin methods and conflicts — dbreunig · 2026-10-27
- Aidan Gomez: LLM probing outside servers was out-of-bounds, not proof of intent — AccBalanced · 2026-10-27