b9982
ggml-org/llama.cppb9982Jul 13, 2026by github-actions[bot]same day as b9981
What changed since b9981
Server now honors per-request reasoning_budget_tokens/message in chat completions, fixing Qwen3 reasoning control.
New features
- Per-request reasoning budget overrides in chat completions are now respected, enabling advanced reasoning models such as Qwen3 to be controlled per call.
Improvements
- Added a unit test in test-chat.cpp that exercises the reasoning-budget parsing path through oaicompat_chat_params_parse with a Qwen3 template.
- Parameter parsing order corrected so caller-supplied values are read before the generic body-copy loop runs.
Bug fixes
- Per-request reasoning_budget_tokens values were silently discarded in favor of the server default (typically -1); the request body value is now used.
- Per-request reasoning_budget_message was likewise ignored and overwritten by the server-level default; it is now read from the request body.
Upgrade notes
No config or migration changes are required; this is a drop-in patch over b9981. If you previously set reasoning_budget_tokens per request (e.g. 0 to suppress thinking) and saw no effect, that override now takes precedence over the server default. Server-level defaults still apply when the request omits the field.
Analyzed Oct 24, 2026
Upgrading?
Release notes summary
Fixes a bug where per-request overrides for `reasoning_budget_tokens` and `reasoning_budget_message` in chat completions were ignored in favor of server defaults, breaking support for advanced reasoning models like Qwen3.
Key Highlights
- Per-request reasoning budget fix
- Qwen3 support
- Server parameter parsing correction
Release Notes (excerpt)
<details open> server: honour per-request reasoning_budget_tokens in chat completions (#23116) * server: honour per-request reasoning_budget_tokens in chat completions The reasoning-budget block in oaicompat_chat_params_parse read only the server-level default (opt.reasoning_budget, typically -1) and the Anthropic-style alias thinking_budget_tokens, but never the canonical reasoning_budget_tokens field from the request body. Because the key…
Published by ggml-org on GitHub.
Read the full release notesRelated Repositories
:electron: Build cross-platform desktop apps with JavaScript, HTML, and CSS
Godot Engine – Multi-platform 2D and 3D game engine
ClickHouse® is a real-time analytics database management system
Self-hosted game stream host for Moonlight.
DuckDB is a high-performance analytical database system