The ggml-org team has released llama.cpp b12019. It updates chat handling for Qwen3-Coder: with reasoning_format set to none, the software can accept two think-block variants and strip the block from the response.
The release also includes ready-to-download builds for macOS, iOS and several Ubuntu systems.
Why it matters
People running Qwen3-Coder through llama.cpp can use the updated handling to avoid think blocks appearing in responses configured to omit reasoning, though the release notes do not specify which applications or setups are affected.
