A new build of ggml-org/llama.cpp, b11782, was published. This repository is an LLM inference project implemented in C/C++, and this build fixed a problem where the number of completion tokens recorded in usage was counted lower than the actual value when the server ran multiple candidates. The release was published as the ggml-org/llama.cpp b11782 release tag on GitHub, and the release date is October 21, 2026. What was touched is the aggregation output that the server records.
The completion token count means the number of tokens the model newly generated, excluding the input in usage. Since usage is the aggregation output where the server records token counts per request, it becomes a yardstick for gauging inference cost and response length. When the server runs multiple candidates, missing some of them leads to underestimating the total generated amount, and the purpose of this fix is to close that gap. The repository description says only one line, 'LLM inference in C/C++', and the current star count is 131868.
However, the b11782 release notes disclose only this counting fix. Whether other changes came along cannot be confirmed from the release record alone. Figures comparing throughput or latency, and license information, are also not included in this material. Because the change touches only aggregation accuracy, its real impact on server performance is hard to judge from this information alone.