GitHub - antirez/ds4: DeepSeek 4 Flash local inference engine for Metal
DeepSeek 4 Flash local inference engine for Metal. Contribute to antirez/ds4 development by creating an account on GitHub.
Welcome! Type "help" for available commands.
$
Summary
DwarfStar 4 (ds4) is a C inference engine by antirez running DeepSeek V4/V4.1 Flash, GLM 5.x, and Qwen3.8 Flash Next on Metal, CUDA, and ROCm.
It uses asymmetric 2-bit quantization and SSD-backed KV cache, exposing a CLI, OpenAI/Anthropic-compatible HTTP APIs, and a native coding agent. Target hardware is 128 GB+ machines; M5 Max benchmarks show ~34 t/s generation at 2K context.
