vLLM v0.32.0 was released on October 18, 2026, with the vLLM project publishing a GitHub release that bundles 814 commits from 363 contributors, 121 of them first-timers. For teams serving open-weight models on their own hardware, the parts that matter are not the new model list: low-precision KV cache support now reaches older GPUs, and several familiar defaults have changed outright, so upgrading without reading the notes can break a working deployment.
A new endpoint for judgment calls, not generation
The release adds a POST /v1/systemone endpoint, enabled by default, that answers structured judgment questions over a given state: which option to choose, whether a claim holds, and how to score a result. Teams previously coaxed these answers out of models with formatting prompts and parsed them by hand; the serving layer now treats them as a first-class interface. A generic hf parser also reads the response_template shipped inside a model's checkpoint, so models carrying their own template no longer need bespoke tool-call and reasoning parsers, removing the fiddliest part of onboarding a new model.
The KV cache trade: half the memory, paid for in first-token latency
Memory is the first wall that long contexts and high concurrency hit, and most of this release's effort went there. NVFP4 KV cache now runs on SM8x, SM12x and SM90 hardware, letting older cards free up memory with low-precision caching. GLM-5.x switches to an FP8 KV cache by default on SM100, halving per-card KV memory while adding 2.3 to 5.5 percent throughput at a 7 to 11 percent cost in time to first token; on SM90, fp8_ds_mla lifts GLM-5.3-Flash output throughput by up to 31 percent. The trade is stated plainly: memory comes back, and users wait slightly longer for the first token. Low-level inference optimization has been a running theme this year, as in the vLLM optimization that let Kimi K3 beat a B300 on AMD hardware.
New models, short-request throughput and security fixes
Newly supported models include Google's omni-modal embedding model EmbeddingGemma 2, MiniCPM-V 4.7 and NVIDIA's streaming speech recognition model Nemotron 3.5 ASR. On the serving side, requests across multiple API processes now go to the least loaded one, which the project's benchmark credits with 43 to 85 percent more requests per second on short requests — under a very short 32-input-token, 1-output-token workload, so production traffic should not expect the same figure. The release also fixes three tracked vulnerabilities covering chat-template resource exhaustion, a tokenizer parameter and peer-to-peer communication, restricts cross-origin access by default instead of allowing every origin, and limits untrusted images to raster formats.
Check these three breaking changes before upgrading
The breaking changes are more likely to bite than the new features are to delight. Python 3.10 support is dropped, so deployments still on it will fail the upgrade outright. The tighter cross-origin default will break browser front ends that call the inference API directly until allowed origins are configured explicitly. Cascade attention is removed, and the supported transformers range narrows to between 5.16.1 and 5.20.0. The safer order for production is to verify those three points in a test environment first, then measure whether the new KV cache defaults actually help on your own traffic, because the first-token latency cost only shows its true size under your own request distribution.
