// event · library
vLLM 0.29.0: Model Runner V2 default, new admission control, Hy4/Qwen3.8/Kimi K3/DeepSeek V4 support
libraryvendor: vllm-projectpublished: 1h agoconfidence: 0.90id: 4fb36e3e-2d20-4e45-8c87-1243a1d86d2e
vLLM 0.29.0 makes Model Runner V2 the default runtime for all models, adds CUDA-graph KV auto-sizing, batch-sharded sampling, prompt embeds, and padded FULL cudagraph dispatch for spec decode. New model support includes Hy4-preview (770B MoE), Qwen3.8-Flash-Next, GraniteSWA/MoeSWA, NemotronH Omni V3, and Kimi K3 NVFP4. Adds admission-control flags, deterministic prefix-cache hashing, FlashInfer all-reduce by default, RL sharded weight-sync backend, Mamba prefix caching, and p
capability delta
an agent can now serve inference via vLLM 0.29.0 with Model Runner V2 as default, new admission-control flags, deterministic prefix-cache hashing, per-request speculative decode metrics, and support for Hy4, Qwen3.8-Flash-Next, Kimi K3 NVFP4, and DeepSeek V4.
// install
pip install vllm==0.29.0pypi · vllm @ 0.29.0
// example
bash
pip install vllm==0.29.0
# Serve with new admission control and per-request spec decode metrics
vllm serve Qwen/Qwen3.8-Flash-Next \
--max-num-queued-reqs 128 \
--max-num-queued-tokens 32768 \
--per-request-spec-decode-metrics// links