As of 2026-10-25 KST, ggml-org/llama.cpp had published 13 builds (b11857~b11877). This batch contains releases only, and the changes fell roughly into two branches: one was the expansion of backend kernels and model support, and the other was safeguards for the server and RPC paths.
OpenCL gained matrix/vector operations for Q2_K quantization, tiled matrix multiplication, and GEMM/GEMV paths for Adreno (b11873). Top-k selection and a lightning indexer followed (b11874), and Hexagon added per-KV-head team partitioning to multi-core flash attention so that remaining cores can still participate (b11868). Vulkan fixed a bug where warp tiles overwrote the columns of neighboring tiles on 16-way subgroup devices (Mali) (b11876), and increased 3D convolution tile sizes, limited to the coopmat2 path (b11877). b11857 made it possible to skip a computation by specifying index -1 in an indirect matrix multiplication, and b11875 generalized Hexagon's dimension reduction and reordering so that it applies uniformly to unary, binary, and matrix multiplication operations.
On the model side, the consistency of the mixture-of-experts (MoE) path was refined. b11858 rejects tensor operation sharding on architectures where max reduction wraps entire rows and therefore cannot be computed with axis-based sharding, and b11861 fixed a problem where MiniMax M3's per-expert NVFP4 scales and per-tensor MXFP8 scales were not being passed into the computation graph. On the server side, b11871 rejects requests with a 400 error when the nesting depth of the request JSON exceeds 128 levels, blocking a path where the server could die from a stack overflow, and b11865 added bounds checking to the RPC receive path to block reads outside the buffer. b11867 refactored the server execution entry point into a structure with role separation, and b11860 corrected the verification procedure for context tests that had been failing at large batch sizes.
However, this batch did not come with performance measurements. Without numbers, it is hard to judge how much the Q2_K kernel tuning for Adreno and Hexagon's team partitioning differ in actual throughput. If your environment picks up individual builds right away, it is worth checking first the b11873 Q2_K path and the b11871 nesting depth limit, the items that change behavior.