ggml-org/llama.cpp b12000 is a release published on 2026-10-30. The repository description presents the project as C/C++ based LLM inference. b12000 adds OCP FP8 E4M3 quantization support. The release notes say it fixed Python type checking, made FP8 conversion reflect the input scale, and corrected a problem where the result was wrongly converted to NVFP4. The source of the changes is the b12000 release description of the repository.
This release bundles the new quantization support together with conversion related fixes. Specifically, OCP FP8 E4M3 was added to the supported targets, the FP8 conversion was changed so that the input scale is reflected, and the problem where the converted result wrongly came out as NVFP4 was fixed. The Python type checking fix is included as well. What can be confirmed is therefore the addition of a supported format and the correction of conversion and checking behavior, while the release notes do not give the specific paths where each change applies.
The provided release information contains no supported hardware or backends, and no performance or accuracy comparison results. The effect of reflecting the input scale on actual results is not given as a number either. The place where the Python type checking fix applies and the conditions under which the NVFP4 conversion problem occurred are not described either. So what b12000 confirms is the change items themselves, and the scope and performance of the new support path are hard to judge from this information alone.