Salvatore Sanfilippo, the creator of Redis known as antirez, has released DwarfStar 4, a tiny C inference engine that runs frontier-scale AI models on a single laptop with no cloud needed.

DS4 squeezes DeepSeek V4 Flash, a 284-billion-parameter model with only 13 billion active parameters per token, into about 76GB of memory. On a MacBook Pro M3 Max it reaches 26.7 tokens per second, which is interactive speed for a model this size.

The trick is asymmetric quantization. The expert layers are compressed to 2-bit while the routing and shared layers stay at full 8-bit precision, because an error in the router breaks everything downstream.

DS4 is MIT-licensed and deliberately narrow. It ships its own validated weights, plus a coding agent and tool calling. The project crossed 13,000 GitHub stars in a month and caught the attention of llama.cpp’s author.

The engine matters for privacy and cost. Builders paying per-token API bills now have a serious option for running a large model on hardware they own.