# Private CPU runtime

The model requires the accompanying patch; GGUF weights alone do not add support to a stock runtime.

```bash
git clone https://github.com/AtomicBot-ai/atomic-llama-cpp-turboquant.git runtime
cd runtime
git checkout cd560939087c95b93a1f30a95603d6b079436952
git apply ../runtime.patch
tar -xzf ../runtime-new-files.tar.gz
cmake -S . -B build -DGGML_CUDA=OFF -DGGML_METAL=OFF -DGGML_NATIVE=ON -DLLAMA_BUILD_SERVER=OFF -DCMAKE_BUILD_TYPE=Release
cmake --build build --target llama-mtmd-cli -j 16
```

Keep all six files of a variant together. A CPU image example, from the model directory:

```bash
runtime/build/bin/llama-mtmd-cli \
  -m Ling-3.0-flash-VL-AD-Q4_K_M-00001-of-00006.gguf \
  --mmproj mmproj-Ling-3.0-flash-VL-F32.gguf \
  --image example.png -p "Describe this image." --jinja -ngl 0
```

The patch includes model conversion, vision bridge, mRoPE/position handling, and the multi-image CLI batch lifetime fix. The bundled files are source patches, not portable prebuilt binaries. Native compilation requires the usual C++ compiler and CMake build dependencies.