The vLLM project released v0.32.0 on October 18, 2026, spanning 814 commits from 363 contributors, 121 of them new. Read the release notes.
Low-precision KV cache
- NVFP4 KV cache now runs on SM8x and SM12x, SM90, and with XQA decode on SM12x, plus on MiniMax-M3 sparse attention on SM100.
- GLM-5.x defaults to an FP8 KV cache on SM100, halving KV memory per rank with 2.3–5.5% more throughput at a 7–11% time-to-first-token cost in the cited benchmark; on SM90, the fp8_ds_mla path gives GLM-5.3-Flash up to 31% more output throughput.
- ROCm gains a 4-bit ultraquant KV cache with roughly twice the density of FP8.
Structured decisions and parsers
A new POST /v1/systemone endpoint answers choice, noul and score questions over structured state, is served by default, and also runs on Cloudflare clef-omni, LiquidAI d1-3B and DiffusionGemma. A new hf tool and reasoning parser reads the checkpoint's response template, so models shipping it need no model-specific parser, in both the Python and Rust frontends.
Model Runner V2 and serving speed
Model Runner V2 adds a GPU n-gram speculator with variable-length drafts, Elastic expert parallelism, pipeline parallelism with prefill context parallelism for MLA models, and a stock torch.compile mode; it is now the default on ROCm. Startup and serving improvements include precompiled FlashInfer kernels via vllm download-kernels, warm autotune tables across restarts, a shared-memory tensor arena for multimodal workers, and least-loaded API-worker routing that the benchmark reports as 43–85% more requests per second on short requests.
Model support and security
New model support includes EmbeddingGemma 2 omni-modal embeddings, MiniCPM-V 4.7 and Nemotron 3.5 ASR, plus video input in the Transformers backend and GLM-5.3 sparse-MLA index reuse (3.5–3.9x on that kernel). The release also fixes three disclosed vulnerabilities: chat template resource exhaustion, tokenizer max_token_id handling, and a P2P ZMQ socket issue.
Performance figures above are from the project's pull-request benchmarks, not independent tests; our verification guide describes how we treat them.