Ds4

DwarfStar (ds4) is a small native inference engine built first for DeepSeek V4 Flash, with support for GLM 5.2 and, on very high-memory machines, DeepSeek V4 PRO. It is self-contained and deliberately narrow rather than a general GGUF runner: model loading, prompt rendering, tool calls, KV state, the HTTP server, and the coding agent are built and tested together. Metal is the primary backend, targeting Macs with 96 GB or more (smaller machines can stream from SSD), alongside NVIDIA CUDA and AMD ROCm. The repo also ships tooling for GGUF conversion, imatrix, and quality and speed measurement.