AI Briefing
KO

llama.cpp Confirms reasoning_effort Bug

·2026.08.15 06:47

Key point

A bug has been reported where llama.cpp's llama-server fails to pass the reasoning_effort parameter to the chat template.

Details

When processing the reasoning_effort parameter, llama-server recognizes the "none" value to disable thinking capabilities, but ignores other settings such as low, medium, high and fails to pass them to the chat template.

As a result, for models like Qwen3.8-27B that use Jinja templates to inject different reasoning instructions based on the reasoning_effort value, the model always operates with the default value (e.g., xhigh) regardless of the user's setting.

Key Impacts:

  • No Standard API Control: Clients using the standard reasoning_effort field, such as the OpenAI SDK or VSCode Copilot, cannot control the model's reasoning intensity.
  • Limitations of Workarounds: Only clients that directly manipulate chat_template_kwargs can temporarily mitigate the issue.

Solution: A code fix has been proposed to include non-"none" values in chat_template_kwargs when they are provided.

qwenllama-cppreasoning-effortinference-serverbug-report
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Redditllama.cpp Optimizes Metal for Qwen3.8Hacker NewsApple Releases LensVLM-9B for Processing Compressed Text ImagesTLDR AIQwen Releases 'RecreationWorld' for Training AI Agents to Rebuild AppsHuggingFace BlogHugging Face Adds GGUF Model Support to TransformersRedditSenseNova-U1.5-8B-MoT Released: Unified Text-to-Image and Understanding Model Initialized from Qwen3Hacker NewsClaude Code Fails to Load AGENTS.md When Telemetry Is Disabled, With Workaround