
llama.cpp has merged a fix for an edge case in Gemma 4 tool calls and added support for two additional weight formats on its Hexagon backend. The changes landed separately on November 14, according to the project's b12388 release and Hexagon pull request.
Tool calls during reasoning
The parser fix addresses a case where a model moves from its thought channel directly to a tool call without the usual closing marker. Previously, the reasoning parser could consume the following tool call. The change recognizes the tool-call boundary; it does not change the model's weights or guarantee that every generated call is valid.
More Hexagon formats
The second change adds IQ2_XXS and NVFP4 weight handling to ggml's Hexagon backend, including optimized paths for different batch sizes. The contributor reports backend-operation tests against CPU results. Those are contributor-reported tests, not independent verification by OpenSourceCoder.com.
Compatibility still depends on the model, quantization, hardware and build. We have not run either change. Our llama.cpp repository profile links to the upstream project and its release history.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- ggml-org/llama.cpp ↗View project details
Original reporting; primary sources reviewed November 14, 2026. No independent hands-on test claimed.



