llamacpp-server-macos

llama.cpp llama-server — macOS (Apple Silicon) static builds, by BatiAI

Self-contained llama-server binaries for macOS, built so that BatiFlow can run on-device multimodal (image + audio) models like Gemma 4 12B that Ollama can't serve yet.

Ollama 0.20 doesn't know the Gemma 4 gemma4uv / gemma4ua vision/audio projectors. llama-server from recent llama.cpp master does. These are the macOS binaries that make that work locally.

Status

✅ Live (2026-06-06) — built on Apple Silicon, verified (/props → vision+audio, image OCR), zero external deps. arm64 only (universal2 not needed — all targets are Apple Silicon).

Files

• llama.cpp commit: 308f61c31f083251ce8150f10b9ef97679b500b5
• sha256 (server): 75cb1b97dff5de3ed3dbc3b79bbbda3cc33cb20f2e9287c3d0b685c46ad4c7db
• deps: macOS base frameworks only (Metal/Accelerate/Foundation/libSystem/libc++/libobjc) — no homebrew/openssl/dylib

Build pin (reproducible)

git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
git checkout 308f61c          # gemma4 projector merged (gemma4v/uv/a/ua clip graphs)
cmake -B build -DGGML_METAL=ON -DLLAMA_CURL=OFF \
  -DBUILD_SHARED_LIBS=OFF \         # static single file (no dylib deps)
  -DLLAMA_OPENSSL=OFF \             # drop homebrew openssl@3 dep (crashes on Macs without homebrew)
  -DGGML_METAL_EMBED_LIBRARY=ON \   # embed Metal shaders (no separate .metal file → true single binary)
  -DCMAKE_BUILD_TYPE=Release
cmake --build build --target llama-server llama-mtmd-cli -j
# static check: `otool -L build/bin/llama-server | grep -iE 'libllama|libggml|libmtmd|openssl|ssl'` → empty

Usage (Gemma 4 12B multimodal)

llama-server-macos-arm64 \
  -m gemma-4-12B-it-Q4_K_M.gguf \
  --mmproj mmproj-google-gemma-4-12B-it-BF16.gguf \
  --host 127.0.0.1 --port 8899 -ngl 99 -c 8192 --jinja
# GET /props  → {"vision": true, "audio": true}
# POST /v1/chat/completions  with image_url / input_audio (OpenAI-compatible)

GGUF + mmproj: batiai/gemma-4-12B-it-GGUF (filenames are pinned/stable).

Scope

• ✅ Image (OCR, visual reasoning) · ✅ Audio (speech, experimental) · ❌ Video (no gemma4 video projector in llama.cpp yet)
• --jinja required. Gemma 4 is a reasoning model → give generous max_tokens (answer in reasoning_content + content).
• Model-agnostic: this llama-server works with any GGUF, not only Gemma 4.

Built and hosted by BatiAI for BatiFlow.