llama.cpp is a large language model inference engine written in C/C++. During one day on October 21, 2026 Korean time, 26 builds from b11757 through b11796 were published, and there were no separately counted commit bundles. The bundled builds had three themes: correcting operations in accelerator backends such as CUDA and HIP, OpenCL, SYCL, and Metal; expanding server features; and fixing multimodal processing.
On the backend side, the fixes included CUDA graph cleanup (b11757), a reduction error for tile widths not divisible by 64 (b11760), NEON fp32→fp16 conversion selection (b11761), an OpenCL matrix multiplication variant for small batches plus the addition of dp4a (b11763, b11789), a fix for SYCL flash attention conflicts (b11788), a memory leak in the Metal GPU capture path (b11791), and NEON vectorization of Hexagon repacking (b11790). The server gained a real-time transcription API and a WebSocket proxy (b11780), POST /tts for speech synthesis (b11781), logprobs return on the speculative decoding path (b11776), passthrough of sampling settings for /v1/messages (b11795), and a fix for propagating client connection termination (b11769). On the multimodal side, LFM2 tiling parameters are now read from GGUF metadata (b11767), and tile truncation and batch composition were matched to the Python transformer implementation (b11774). The changes also included scaling MoE down projections per expert token (b11785) and an optimization of vocabulary loading (b11793).
Still, the limits of judgment come from the numbers. These builds are mainly correctness fixes, yet the release notes present no figures for the magnitude of speed improvements or for quality metrics. How much faster the OpenCL, SYCL, and Hexagon paths actually are cannot be known from publicly available information. Also, because a single day's worth of changes brings both new APIs such as transcription and /tts and judge model support (H2O-Lightning-4B, b11773), what to try first has to be decided by weighing the purpose of each build.