---
format: "aidr-story-markdown/v1"
id: "1e78eb133dbbef2d344521cec89c32e5024c9e197a1b030b9d01f9b2f15a1066"
canonical_url: "https://aidr.today/1e78eb13?lang=en"
title: "Fixing Tool-Call Accuracy in vLLM"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-26T12:59:48.000Z"
category: "Frameworks"
topics: ["deepseek","llm","framework"]
source_urls: ["https://www.ubicloud.com/blog/fixing-tool-call-accuracy-in-vllm","https://news.ycombinator.com/item?id=50242870"]
summary: "A vLLM fork achieved 100% tool call schema accuracy with Kimi and DeepSeek models, and vLLM 0.31's constrained decoding support now enables the same result with newer versions of those models."
---

# Fixing Tool\-Call Accuracy in vLLM

> [Open the canonical story](<https://aidr.today/1e78eb13?lang=en>)

**Published:** 2026-10-26T12:59:48.000Z
**Category:** Frameworks
**Topics:** deepseek, llm, framework

## Summary

A vLLM fork achieved 100% tool call schema accuracy with Kimi and DeepSeek models, and vLLM 0\.31's constrained decoding support now enables the same result with newer versions of those models\.

## Sources

- [Story source](<https://www.ubicloud.com/blog/fixing-tool-call-accuracy-in-vllm>)
- [Discussion](<https://news.ycombinator.com/item?id=50242870>)