Redis creator ships DwarfStar 4, a deliberately narrow local LLM engine

Salvatore Sanfilippo (antirez), creator of Redis, released DwarfStar 4 (ds4), an MIT-licensed local inference engine. It runs only project-built GGUF files for DeepSeek V4/V4.1 Flash, GLM 5.x and Qwen 3.8 Flash Next on Mac (Metal), CUDA and ROCm. On an M5 Max 128GB it hits 39.4 t/s generation at 2K context and 27.6 t/s at 65K.

• ds4 repo reached about 23,000 stars in two weeks
• Supports DeepSeek V4/V4.1 Flash, GLM 5.x, Qwen 3.8 Flash Next
• M5 Max 128GB: 790 t/s prefill, 39.4 t/s generation at 2K
• Asymmetric quantization compresses MoE experts to ~2-bit