vLLM remains one of the most widely used open-source engines for serving large language models, and its recent releases show how quickly production inference is evolving. Version 0.24.0 landed with 571 commits from 256 contributors, and 0.25.0 followed with 558 commits from 232 contributors.
Highlights of 0.24
New model support includes MiniMax-M3 with FP8 and MXFP4 paths and extensive AMD tuning, plus DiffusionGemma. DeepSeek-V4 support continued to mature, with a FlashInfer sparse index cache and prefill planning improvements. Model Runner V2 now handles quantized models by default, and a new Streaming Parser Engine unifies tool-call and reasoning parsing across several model families.
On the serving side, DeepEP v2 was integrated for wide expert parallelism, disaggregated serving gained KV transfer from prefill to decode over NIXL, and the Rust frontend added API-key authentication, CORS, tokenization endpoints and request-control routes.
Highlights of 0.25
Version 0.25.0 made Model Runner V2 the standard execution path, added prefix caching for Mamba hybrid models, and announced that the Transformers modeling backend is now as fast as native vLLM implementations, with FP8 MoE support. A quick 0.25.1 patch fixed a startup problem when FFmpeg was missing and guarded a risky kernel fusion.
Before you upgrade
The 0.24 notes include breaking changes. vLLM no longer sets CUDA_VISIBLE_DEVICES internally, so use the new device_ids argument. Transformers v4 and first-generation Qwen/QwenVL models are deprecated, and several older architectures were removed. A security hardening batch also addressed denial-of-service issues in audio and speculative decoding paths.