QuarkStar
A narrow, self-contained inference engine for two specific Qwen models, Qwen3.6-35B-A3B and Qwen3.8-27B, with Vulkan and Metal graph executors tuned for the AMD BC-250 and Apple Silicon, plus a built-in coding agent and HTTP server.
QuarkStar is a small, self-contained native inference engine built specifically for two models, Qwen3.6-35B-A3B and Qwen3.8-27B, rather than a general-purpose GGUF runner. It includes mixture-of-experts-specific Vulkan and Metal graph executors, model loading, prompt rendering, tool calls, KV state handling, an HTTP server and a coding agent, along with tools for GGUF conversion, imatrix generation, and quality and speed measurement.
Its Vulkan backend is generic but its fast path is developed and tuned for the AMD BC-250 GPU running RADV; its Metal backend targets Apple Silicon M1 or newer, with 8 GB Macs expected to use SSD streaming and 16 GB Macs able to keep the default model resident. A CPU backend exists as a portable reference and debugging path rather than for production use. Supported models are chosen for what runs well on 16 to 32 GB of memory, and can be dropped when a better replacement appears.
It is written in C and licensed under MIT, building on llama.cpp and GGML.