> ## Content Index
> Fetch the complete content index at: https://orbiorbi.com/news/llms.txt
> Use this file to discover other available public pages before exploring further.

# vLLM 0.32.0 adds structured decisions and generic parsers
- URL: https://orbiorbi.com/news/en/hot/vllm-0-32-0-structured-decisions-kv-cache-en/
- Published: 2026-10-18T10:23:45.000Z
- Updated: 2026-10-18T23:00:38.000Z
- Description: According to vLLM, 0.32.0 adds the default /v1/systemone endpoint for structured decisions and hf tool and reasoning parsers that read a model's response_template. It also expands low-precision KV cache support.
- Author: Orbis News
- Tags: #orbis-hot, #en

According to vLLM, 0.32.0 adds the default /v1/systemone endpoint for structured decisions and hf tool and reasoning parsers that read a model's response\_template. It also expands low-precision KV cache support.

- The hf parsers support both Python and Rust frontends; models supplying the template need no model-specific parser. [\[1\]](#hot-source-1)
- In the benchmark cited in the notes, default FP8 KV cache for GLM-5.x on SM100 halves KV memory per rank, with 2.3%–5.5% more throughput at a 7%–11% increase in time to first token. [\[1\]](#hot-source-1)

## Sources and originals

- \[1\] vLLM · 2026-10-18 · [Read original](https://github.com/vllm-project/vllm/releases/tag/v0.32.0?ref=orbiorbi.com)  
The release notes describe structured decisions, generic parsers and expanded low-precision KV cache support.

{"version":2,"hot":{"event\_id":"vllm-0-32-0-structured-decisions-kv-cache","group":"open-source","count":1,"heat":20,"verified":"confirmed","firstSeen":"2026-10-18T10:23:45.000Z","lastSeen":"2026-10-18T23:00:30.415Z","sources":\[{"n":"vLLM","publisher":"vLLM","u":"https://github.com/vllm-project/vllm/releases/tag/v0.32.0","say":"The release notes describe structured decisions, generic parsers and expanded low-precision KV cache support.","tier":"T0","at":"2026-10-18T10:23:45.000Z"}\],"signals":\[\],"facts":\["The hf parsers support both Python and Rust frontends; models supplying the template need no model-specific parser.","In the benchmark cited in the notes, default FP8 KV cache for GLM-5.x on SM100 halves KV memory per rank, with 2.3%–5.5% more throughput at a 7%–11% increase in time to first token."\]}}