llama.cpp build in use
Not vendored; prebuilt release downloaded into bin/ (gitignored). Re-fetch with:
curl -sL -o bin/llama.tar.gz https://github.com/ggml-org/llama.cpp/releases/download/b11311/llama-b11311-bin-macos-arm64.tar.gz
tar -xzf bin/llama.tar.gz -C bin && rm bin/llama.tar.gz
ln -sf llama-b11311/llama-server bin/llama-server
Version
- Release tag b11311 (
version: 0.5.0-dev, build 11311, commit f7b384c1e), published 2026-10-01 (UTC), macOS arm64, AppleClang 21. - Nightly pointer release at fetch time was v0.5.0 → b11146; b11311 was the newest tagged binary release.
Capability findings (this build)
| Capability | Supported | How |
|---|---|---|
| Memory-mapped loading | yes | -lm, --load-mode MODE with modes auto (default; mmap unless device unsupported), mmap, mmap+mlock. The old --no-mmap flag is replaced by this unified flag. |
| Partial GPU offload | yes | -ngl, --gpu-layers, --n-gpu-layers N |
| Parallel slots + continuous batching | yes | -np, --parallel N (default -1 = auto), -cb, --cont-batching (on by default; --no-cont-batching to disable) |
| bf16 GGUF | yes | Metal backend dylib shipped (libggml-metal); bf16 GGUF loads verified empirically in item 10 (Metal bf16 path, M4 native bf16). |