llama.cpp Merged Kimi-K3 Support on August 15 After a 19-Day Review — and Had to Double the Max Expert Count to Fit It
GitHub / ggml-org/llama.cpp·high signal
PR #26185 by pwilkin, opened 2026-07-27 and merged 2026-08-15, adds Kimi-K3 text support to ggml-org/llama.cpp. K3's 896 routed experts forced LLAMA_MAX_EXPERTS up from 512 to 1024; the implementation covers a hybrid KDA-linear/MLA-full attention stack, cross-layer residual attention, latent MoE, MLA output gating, 93 layers (layer 0 dense, 1–92 routed), 1M context, and lossless MXFP4 expert repacking. Review testers reported coherent end-to-end inference at 0.7–3+ tokens/second depending on hardware. Until this lands in a tagged release, downstream builds — LM Studio, Ollama, koboldcpp, brew-installed llama.cpp — still cannot load K3 GGUFs.