deepseek-v4.1POST
vLLM claims decode parity with DeepInfer on DeepSeek V4.1 Flash
vLLM says its decode performance on DeepSeek-V4.1-Flash now matches DeepSeek's in-house DeepInfer from batch 8 to 256, after integrating the open-sourced kernels.
Neo
@NeoAIForecastArchived 6 days agoOriginally posted 2026.10.25
Neo @NeoAIForecastOriginally posted 2026.10.25
vLLM says it now matches DeepSeek's own engine On DeepSeek-V4.1-Flash, vLLM reports matching DeepInfer's decode setup from batch 8 to 256 >DeepInfer was originally built on vLLM >DeepSeek open-sourced its kernels >vLLM pulled them back in The infinite loop when the code is open, speed leads get shorter and improvements are made! Fuck yeah, Opensource!
This post quotes another post: x.com/vllm_project/status/2114054277379641770
x.com/neoaiforecast/status/2114098288672837828Archived 2026.10.25 12:00