llama.cpp merges Kimi-K3, doubles expert limit

llama.cpp merged Kimi-K3 text support on August 15, closing a pull request contributor pwilkin opened 19 days earlier, per PR #26185.

K3 ships with 896 routed experts, more than double the 512-expert ceiling llama.cpp's code had hard-coded. The merge forced that limit up to 1024, so anyone running local mixture-of-experts models on llama.cpp now inherits the higher cap, not just Kimi-K3 users.

The PR pairs linear KDA attention with full MLA attention, then layers in cross-layer residual attention and latent MoE routing. MLA output gating spans 93 layers, with a 1M-token context window and lossless MXFP4 repacking for the expert weights.

None of that runs anywhere convenient yet. Testers on the PR reported coherent output at 0.7 to 3-plus tokens per second. The code hasn't reached a tagged llama.cpp release, so LM Studio, Ollama, koboldcpp, and the brew-installed build still can't load Kimi-K3 GGUF files.

The number worth watching is 1024, not K3. llama.cpp's expert cap was hard-coded at 512 until one model broke it, and bigger sparse models are becoming the norm. Builders who want K3 locally can wait for a tagged release instead of compiling the branch. 0.7 to 3 tokens per second is the price for early access.