deepseek-v4.1IMAGE
Maya update: DeepSeek V4.1 Flash 1017 prefill up nearly 3x on 3x V100
On the same 3x Tesla V100 32GB setup with the same DeepSeek V4.1 Flash 1017 quant, a Maya update lifted prefill from about 420 to about 1400 tok/s without hurting decode.
AJ
@ItsmeAjayKVArchived 8 days agoOriginally posted 2026.10.24
AJ @ItsmeAjayKVOriginally posted 2026.10.24
Heads up! Maya just got faster ! On same 3x Tesla V100 32GB + DDR4 system running same DeepSeek-v4.1-Flash-1017 quant just got a whole lot faster. Prefill is improved by almost 3x, before: ~420 tok/s after: ~1400 tok/s That's huge, all while decode is not going down !! @PeasantSmith is moving fast with Maya, much appreciated.
This post quotes another post: x.com/peasantsmith/status/2113730024125927571
x.com/itsmeajaykv/status/2113773463148081488Archived 2026.10.24 12:06