The b11772 release of ggml-org/llama.cpp was published on the GitHub releases page on October 21, 2026. This change fixes a bug in gemma 4's FFN activation function in mtmd, and it is the only change listed in the published release notes. An FFN (feed-forward network) is a set of neural network layers that comes after the attention layer in a transformer model, and the activation function is the part that makes the output of that operation nonlinear.
As the repository description states, llama.cpp is a project that implements LLM inference in C/C++, and it currently has 131,859 stars. The star count only indicates the size of the repository and does not speak for inference quality. The subject of this release is not the project as a whole but the processing path of the single model gemma 4, and the nature of the change is a defect fix rather than a feature expansion. The published notes do not reveal which inputs and models mtmd is responsible for, so how far this fix reaches requires confirmation from the repository side.
For now, the public information goes only as far as the fact that something was fixed. Which inputs revealed the problem, what the activation function had been wrongly set to, and how much the results differed are not given. No test results or performance measurements are presented either. So it is safe to read this change as a defect fix that restores the accuracy of the mtmd processing path that handled gemma 4. There is no evidence that other models or other modules were affected.