The C/C++-based LLM inference project ggml-org/llama.cpp released b12202 on November 8, 2026. This release changes some operations in the reference implementations that were not computed in 32-bit floating point to bfloat16 precision. The changes were published on the release page. This is not a blanket change to the entire inference path, but an operation-by-operation precision adjustment as specified in the release summary.
The targets are the down projection and output projection of the GLM4 family, the visual matrix operation of MiMoVL, and the convolution of the Qwen3-TTS speaker encoder. In other words, this change adjusts precision by designating different operations in three model families. The release summary names the two projection operations of GLM4, the visual matrix operation of MiMoVL, and the speaker encoder convolution of Qwen3-TTS as the respective change targets. The core of the change is that these operations are now computed in bfloat16 precision.
However, the precision change was not applied to every operation. The DeepSeek V4/V4.1 router and the MiniMax-M3 indexer keep 32-bit. The published change details clarify which operations changed and what was kept, but do not state the reason for the precision split or its effect on speed and accuracy. Therefore, what can be confirmed in this release is a change in the precision assignment of the designated operations, and its effect is difficult to judge from public information alone.