The b11870 release of ggml-org/llama.cpp officially supports the DeepSeek-V4 Flash model. The repository description states that it is a C/C++ based LLM inference project. This change also includes an n-gram table format change. An optimization that computes only once when reading compressed key-value entries came along with it. The source of the changes is the ggml-org/llama.cpp b11870 release on GitHub. The release date is October 25, 2026. So the key items confirmed in this version are the new model support and two inference-related changes.
The n-gram table has a changed format, and the compressed key-value handling now has an optimization that reduces the read computation to a single time. However, the provided release summary does not explain the specific structure of the new table format or how it is compatible with the previous format. It also does not present the detailed paths where the compressed key-value optimization applies or the comparison criteria. What can be confirmed here is the kind of the change and the summarized intent. The specific scope of the implementation is difficult to judge from the release summary alone.
The release summary explicitly states DeepSeek-V4 Flash support, but the supported devices or settings are not given. Performance measurements and quality comparisons are also not included in the provided facts. Therefore, there is no basis to read the official support in b11870 as a speed or quality improvement in a specific environment. This release is an update announcing model support and processing changes, while actual performance and compatibility remain outside the scope presented in the summary.