vLLM v0.25.0 launches with Model Runner V2 as default, drops PagedAttention
558 commits from 232 contributors make vLLM faster and more capable.
vLLM v0.25.0 brings major architectural changes, including Model Runner V2 (MRv2) as the default execution path for all dense models, building on quantized-model support from the previous release. The legacy PagedAttention implementation has been removed entirely, as V1/MRv2 backends are now the standard. The Transformers modeling backend is now as fast as native vLLM, gaining FP8 MoE support, CUDA graph fixes, and migration of GPTBigCode/Starcoder2 and RoBERTa. New models added to the zoo include LLaVA-OneVision-2, Unlimited OCR, MOSS-Transcribe-Diarize, and Hy3, along with GLM-5/DeepSeek-V3.2 and MiniMax-M3 with pipeline parallelism and NVFP4 support.
On the speculative decoding front, universal support for heterogeneous vocabularies (TLI) is introduced, along with new drafters DSpark and DFlash. The Rust frontend continues to mature with HTTPS/mTLS, a DP supervisor, and profiler control routes. Engine core improvements include realtime embeddings, prefix caching for Mamba hybrid models, multimodal-prefix bidirectional attention, and dynamic speculative decoding compatible with full CUDA graphs. A new Streaming Parser Engine provides a unified tool-call/reasoning parsing framework with ports for Kimi k2.5/k2.6/k2.7, seed_oss, and DeepSeek V4.
- Model Runner V2 is now the default for all dense models, replacing legacy PagedAttention with V1/MRv2 backends.
- Transformers backend now matches native vLLM speed with FP8 MoE support and fixes for CUDA graphs and embedding scaling.
- New models include LLaVA-OneVision-2, GLM-5/DeepSeek-V3.2, MiniMax-M3 (with pipeline parallelism), and universal speculative decoding for heterogeneous vocabularies.
Why It Matters
vLLM continues to dominate open-source LLM inference, adding speed, model support, and enterprise features.