The v0.28.0 release of the vLLM open-source project introduces 584 commits from 270 contributors, marking a significant performance push for Kimi-K3. Key optimizations include Decode Context Parallel (DCP) support and fused Flash KDA kernels, delivering a 1.5x to 3x kernel-level speedup via combined all-gathers. The release adds sparse MLA support for DeepSeek V4, enabling end-to-end decode and speculative decoding, alongside AMD Quark NVFP4 support.

Speculative decoding advances with DFlash2 and confidence-scheduled DSpark verification. Model Runner V2 maturation features E/P/D disaggregation and weight offloading. New defaults raise max batched tokens to 16384 and enable prefix caching for Mamba models.

Breaking changes include the migration of bitsandbytes to an out-of-tree plugin and a Transformers bump to 5.15.0. Release artifacts include Python wheels for CUDA 13.0 and ROCm, plus Docker images across multiple platforms. New model support features Muse Glimmer, Ling 3.0, and Qwen3.5 enhancements.

The update solidifies vLLM's position as a leading framework for large language model serving.