DWARFSTAR 4

Summary from the TLDR Dev newsletter

DwarfStar 4 is a self-contained native inference engine for the DeepSeek V4 Flash model that supports Metal and CUDA backends and is optimized for local performance with features like 2-bit quantization and a compressed KV cache handling up to one million tokens. The engine provides a command-line interface and an OpenAI-compatible server for integrating coding agents on high-end personal hardware.