DeepSeek V4.1 measures its own decode speed at 430-460 tok/s — then the claim falls apart
teortaxesTex · x · 2026-10-15
teortaxesTex shared a screenshot of DeepSeek V4.1 (shown as DeepSeek-V41-Flash High) analyzing its own inference speed during a coding task: the model computed slopes across multiple completions and concluded a true steady-state decode rate of 430-460 tok/s after removing TTFT, noting a 1-2s scheduling ramp; the session showed a 99% cache hit rate.
The author later walked it back: after checking the fingerprint (aeb56401...), the actual speed excluding TTFT is only 240 tok/s — "false alarm, I was high on hopium." Fun moment of a model measuring itself, but the headline speed claim doesn't hold.
Related event: DeepSeek V4.1 Rumor Debunked After Fingerprint Check(2 posts)→
More from Fun
- Jensen Huang recalls delivering NVIDIA's first AI supercomputer to OpenAI after only Musk ordered — elonmusk · 2026-10-15
- Dev puts a Muse AI agent on the wall with Pi 5 and laser projector — alexandr_wang · 2026-10-15
- Claude wrote its own Direct3D 9 and x86 emulators to run kkrieger in a browser — austinc3301 · 2026-10-15
- Figure 02 humanoid crosses obstacles blind, powered by an RL controller — jonstephens85 · 2026-10-15
- Claude Code quota ran out, GPT-6 Astra turned a Rick & Morty frame into a game — AlchainHust · 2026-10-15
- Claude-generated India startup exits by university: IIT Delhi leads with $80B — geoffwolfe · 2026-10-15