ggml-org/llama.cpp b11700 is a release that added the conversion and support paths for handling decision (judgment) models of the IBM Granite Switch family. It is a build published on October 16, 2026, and llama.cpp is a project that performs large language model inference in C and C++. This change is a small feature addition for decision model support.
The final build contains the converter's Granite Switch decision converter, the corresponding model enumeration in common, and the decision support branch in the server. The template example follows how ibm-granite/granite-switch-4.1-8b uses guardian together with decision. It also fixes that supporting decision also means supporting text output, and removes mentions of decision.chat from the server README. The three changes cover conversion, common model identification, and the server-side branch respectively.
During review, the draft's decision.chat template, the dual chat and decision paths in common, and the server's decision_chat flag routing were all reverted. So the final change should not be read as a separate chat template or as an implementation with two paths. The provided release information presents no performance measurements or scope of supported models. What can be confirmed is the addition of Granite Switch decision support code; its actual applicability and performance are hard to judge from this record alone.