Show HN: Qwen3.6-35B-A3B on a 16 GB M1 Pro with SSD-streamed MoE (github.com)

A new project called DwarfStar has been announced, offering a specialized local inference engine designed for large AI models that typically exceed the memory limits of standard devices, specifically targeting 16–64 GB Apple Silicon systems. This fork of the antirez/ds4 project, led by developer andrea borio, focuses on adaptive SSD streaming and Metal optimization, allowing developers to run large models like Qwen3.6-35B-A3B directly from SSD storage while maintaining acceptable performance. This development is significant for the AI/ML community as it enhances accessibility to advanced model inference on widely used consumer hardware, providing a pathway for experimentation without requiring high-end server setups. With features like native model loading, prompt rendering, and an adaptive cache mechanism, DwarfStar optimizes memory usage by intelligently managing resources between RAM and SSD. The project emphasizes not competing with existing solutions but enhancing capabilities within the Apple ecosystem while ensuring all improvements are shared with the upstream repository to benefit the broader development community.