deepseek-v4.1POST

SGLang low-latency mode on V4.1 Flash: 928 tok/s on GB300, 516 on MI355X

SGLang's low-latency mode with DSpark runs DeepSeek V4.1 Flash at 671→928 tok/s on GB300 (1.38x) and 332→516 tok/s on MI355X (1.56x), with the software open.

Neo

@NeoAIForecastArchived 11 days agoOriginally posted 2026.10.20

Neo @NeoAIForecastOriginally posted 2026.10.20

Open original on X

SGLang Low-Latency Mode, built with NVIDIA and AMD, running DeepSeek V4.1 Flash with DSpark: GB300: 671 to 928 tok/s (1.38x) MI355X: 332 to 516 tok/s (1.56x) The speedup is software, and this software is open

This post quotes another post: “SGLang low-latency mode: DeepSeek V4.1 Flash hits 928 tok/s P90 on GB300”

x.com/neoaiforecast/status/2112367263026430372Archived 2026.10.20 12:02