Antirez adds distributed inference to DwarfStar

DwarfStar is a masterclass in "intentionally narrow" systems engineering, prioritizing vertical integration for DeepSeek V4 over the general-purpose flexibility of runners like llama.cpp. Serial distribution via layer sharding allows users to pool the VRAM of multiple devices (e.g., two Mac Studios) to run models that would otherwise be impossible locally. The disk-backed KV cache architecture elegantly sidesteps RAM limits, allowing for context windows up to 1 million tokens with persistent session state. Native support for asymmetric 2-bit quantization makes the massive PRO model surprisingly accessible, achieving ~13 tokens per second on prosumer hardware. Integrated activation steering and direct agent-in-engine execution suggest a future where inference engines are active participants in the reasoning loop, not just passive token generators. Built in just weeks with significant GPT-5.5 assistance, the project serves as a high-profile validation of AI-accelerated systems programming.