Dev Radar
Support
LiveUpdated 2026-11-03 22:02 UTC

3 days of work: DeepSeek V4.1 Flash answers 2.9x faster on my rig.

3 days of work: DeepSeek V4.1 Flash answers 2.9x faster on my rig. Prefill, stock Maya v1.1.51 vs my build: 16K 1029…

This is a dev post classified by AGK Intelligence as AI dev tools (a tutorial), kept by the Dev Radar because it carries real work, not commentary.

3 days of work: DeepSeek V4.1 Flash answers 2.9x faster on my rig. Prefill, stock Maya v1.1.51 vs my build: 16K 1029 -> 1262 tok/s 116K 1753 -> 2651 tok/s 272K 1406 -> 2888 tok/s, 2.1x First token on 272K: 195 s -> 95 s. On 116K: 66 s -> 44 s. Decode stays about 70 tok/s on both. The 2.9x is first token on replays of my real Claude Code and Codex sessions, 4.7 s -> 1.6 s against stock v1.1.46. 4x 3090, PCIe 3.0, no NVLink. Stock on 272K is unstable, it pulls about 60 GB of experts per card from page cache and its 2nd run did only 622 tok/s. Above is its better run. Below: what I tried, w

Posted by Alexey Fateev (9.7k followers) 3 h ago · 30 likes · 1.5k views · view the original post on X. Kept by the Dev Radar as AI dev tools.

More dev work like this

Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.3k posts from 5k X accounts over the last 21 days, 9.2k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-03 22:02 UTC. Full method.