vllm v0.31.0 released with DeepSeek-V4.1-Flash and preload daemon
vllm v0.31.0 adds DeepSeek-V4.1-Flash performance paths, a 'vllm preload' weight-cache daemon, Model Runner V2 speculative decoding, large-scale serving updates, and security and breaking-change notes.
What changed
DeepSeek-V4.1-Flash SM100 default; 'vllm preload' weight-cache daemon; Model Runner V2 speculative decoding; MoonEP all2all backend; multiple security and breaking changes.
Why it matters
Developers and operators running GPU serving get faster restarts via the preload weight-cache, new SM100/SM103 kernels and MoE/MRV2 fixes that reduce OOMs and improve large-scale serving.
Who should care
DevelopersResearchersEnterprises
Worth trying
Yes — pip and docker artifacts published for CUDA 13.0, ROCm, CPU and XPU.