skip to content
The Weighted Average

Wire

llama.cpp adds a Qwen3 tool-call parser

llama.cpp build b10227 added a specialized Qwen3 parser for tagged reasoning and tool calls, then published 25 downloadable artifacts spanning macOS, Linux, Android, Windows, iOS, and the web UI. The August 2 release says the parser now tolerates omitted <tool_call> wrappers and Qwen3-Coder function delimiters, making local agent output less brittle at the boundary between reasoning and execution. For teams weighing open-weight control against hosted convenience, parser compatibility belongs in the deployment bill: a model that emits the right call in the wrong envelope still breaks the agent loop.