ds4

DwarfStar is a specialized inference engine designed primarily for optimized performance with DeepSeek V4 Flash and GLM 5.2, enabling efficient model execution on high-memory systems and consumer hardware. Key features include support for various backends such as Metal and NVIDIA CUDA for multi-GPU setups, SSD streaming for devices with limited RAM, and the ability to implement both tensor and pipeline parallelism for enhanced model performance and capacity. The tool is streamlined for specific model types and integrates essential components like model loading and tool calls within a single package.