vLLM v0.28.0 ships Kimi-K3 perf push, full DeepSeek-V4 sparse-MLA, Model Runner V2 maturity, and tiered KV-cache offloading
vLLM Project
Released Aug 28, 2026. 584 commits from 270 contributors. Headline: Kimi-K3 decode context parallel, fused FlashKDA kernels, SiTU activation for MegaMoE, GEMM-RS sequence parallelism, ~60% better DSpark TTFT via adaptive speculative budget; ROCm support added. DeepSeek-V4: sparse MLA end-to-end for decode, MTP, DSpark; AMD Quark NVFP4; reasoning-effort mappings. Speculative: DFlash2 with local convolution, DSpark confidence scheduling. Model Runner V2: E/P/D disaggregation, weight offloading, multi-layer MTP KV cache, encoder CUDA graphs, attention-free models. Tiered KV-cache offloading with disk tier, out-of-tree secondary-tier managers, partial load results, tiering metrics. Rust/gRPC frontend gains multimodal image inference and protobuf schemas on Buf. New defaults: max_num_batched_tokens 8192 to 16384, prefix caching on for Mamba, Blackwell CUDA graph capture 1024. Breaking: bitsandbytes moved OOT plugin, Transformers 5.15.0 baseline, calculate_kv_scales and override_attention_dtype removed. Online MXFP4/NVFP4; MLA CPU backend for DeepSeek-V2/V3.
Why it matters
Largest vLLM release of August and the first where the new Model Runner V2 stack, a real tiered KV offload system, and Kimi-K3 day-one perf are all stable together; DeepSeek-V4 going end-to-end on sparse MLA + DSpark makes vLLM the de-facto inference runtime for the latest Chinese-MoE tier.
Importance: 4/5
flagship release; official confirmation