// event · library

vLLM 0.28.0 ships Kimi-K3 + DeepSeek V4 support, DFlash2 spec decoding, tiered KV offload, Rust gRPC frontend

library

vendor: vllm-projectpublished: 2d agoconfidence: 0.95id: 27380ed0-9b94-4834-ab45-a2baa3505e16

vLLM 0.28.0 lands 584 commits from 270 contributors, headlined by a Kimi-K3 optimization push (DCP, fused FlashKDA kernels, GEMM-RS, adaptive spec budget), end-to-end DeepSeek V4 sparse MLA with DSpark, new DFlash2 speculative decoding, Model Runner V2 E/P/D disaggregation, tiered KV cache with disk offloading, a Rust gRPC frontend with multimodal image inference, higher default batched-token and CUDA-graph caps, plus a bitsandbytes plugin split-out.

capability delta

an agent can now serve Kimi-K3, DeepSeek V4, and new MoE/multimodal models via vLLM 0.28 with faster speculative decoding (DFlash2, DSpark), tiered KV cache with disk offload, Model Runner V2 E/P/D disaggregation, and a Rust gRPC frontend.

// install

pip install vllm==0.28.0

pypi · vllm @ 0.28.0

// example

bash

pip install vllm==0.28.0

# Serve a model with the new defaults (max_num_batched_tokens=16384)
vllm serve deepseek-ai/DeepSeek-V4 \
  --tensor-parallel-size 8 \
  --enable-prefix-caching \
  --speculative-model-config '{"method":"dspark"}'

// links

vllminferenceservingspeculative-decodingkv-cachemoekimi-k3deepseek-v4rocmgrpc

source: github_release · https://github.com/vllm-project/vllm/releases/tag/v0.28.0