Release history
vLLM 0.28.0
584 commits from 270
vLLM 0.27.1
VLLM adds support for quantized DSpark Markov heads, enabling efficient inference with compressed attention mechanisms
vLLM 0.27.0
561 commits from 242; Review the changelog before upgrading production inference workloads
vLLM 0.26.0
VLLM 0.26.0 adds Inkling model support, DeepSeek-V4 performance optimizations, fp32 generation heads, flexible attention backends, and mature KV offloading with tiered storage
vLLM 0.26.0
411 commits from 212; Check release notes for model and hardware support changes
vLLM 0.25.1
VLLM 0.25.1 fixes FFmpeg import blocking model startup and mixed-dtype allreduce corruption in quantized models
vLLM 0.25.0
558 commits, 232; Compatibility and performance sweep
vLLM 0.24.0
571 commits from 256; No breaking changes called out in summary
vLLM 0.23.0
DeepSeek-V4 hardening, Model Runner V2 expansion to Llama/Mistral, Rust frontend maturation with streaming and LoRA endpoints, Gemma 4 support, Transformers v5 compatibility, multi-tier KV cache offloading
vLLM 0.23.0
Minimax M3 models remain unsupported in this version
vLLM 0.22.1
VLLM 0.22.1 adds Mellum v2 support, enables zentorch-accelerated quantized inference on AMD Zen CPUs, and fixes multi-node Ray serving hangs and model initialization regressions
vLLM 0.22.0
VLLM 0.22.0 hardens DeepSeek V4 with fused kernels and CUDA graphs, advances Model Runner V2 toward default, adds experimental Rust frontend, and enables multi-tier KV cache offloading for extended memory management
vLLM 0.21.0
VLLM 0.21.0 adds KV offload + hybrid memory allocator integration, speculative decoding with thinking budgets, TOKENSPEED_MLA for Blackwell, and requires C++20 and Transformers v5
vLLM 0.20.2
VLLM 0.20.2 fixes DeepSeek V4 sparse attention hangs, KV cache allocation failures, gpt-oss torch.compile compatibility, and Qwen3-VL load issues