This public repository packages a pinned 4-bit conversion of Gemma 4 12B with a self-contained CUDA llama-server entry point, a larger context allocation, and a plain-text retry if constrained generation fails.

All model and runtime files are local to the repository. The entry point does not download packages or model artifacts during evaluation.

The model is distributed under Apache License 2.0. The bundled llama.cpp runtime is distributed under the MIT License. See LICENSE, THIRD_PARTY_NOTICES.md, and runtime/LICENSE.llama.cpp.