What is ds4?

DwarfStar 4, also called ds4, is a small inference engine built to run one specific AI model, DeepSeek V4 Flash, as fast and correctly as possible on your own computer. Unlike general purpose model runners that try to support many different models, this project is narrow on purpose. It does not wrap another runtime and is self contained, handling model loading, prompt formatting, tool calling, and memory state for that one model, along with a server API and a command line interface for coding agents or direct use. The project targets Apple computers with Metal graphics, starting from MacBooks with 96 gigabytes of memory, as well as NVIDIA GPUs using CUDA, with a separate community maintained branch for AMD hardware. It builds on the ideas and code from the well known llama.cpp and GGML projects, and the authors credit that work directly, including keeping some of its licensed code and copyright notices inside this project. The authors explain why they picked this particular model for a dedicated engine. DeepSeek V4 Flash is described as faster than similarly capable dense models because it activates fewer parameters at a time, its internal reasoning text tends to be shorter and scales with how hard the question is, it supports a context window of one million tokens, and its memory state compresses well enough that it can be saved to disk instead of only living in RAM. The README also states this software was built with heavy assistance from an AI coding tool, with humans directing the design, testing, and debugging, and says plainly that people uncomfortable with AI assisted code should not use it. The project is labeled alpha quality, since it has only existed for a short time and inference engines are complex to get right. It only works with specially prepared DeepSeek V4 Flash model files released for this project, not arbitrary GGUF files from elsewhere. Additional documentation covers contributing guidelines, tools for generating and quantizing model files, quality testing against official outputs, and benchmark results, each in its own linked file within the repository. The full README is longer than what was shown.