DwarfStar 4, the MIT-licensed inference engine from Redis creator Salvatore Sanfilippo (antirez), is drawing fresh attention after a new demo showed it running DeepSeek V4.1 Flash with aggressive compression on consumer hardware, AI Weekly reported.

The engine’s trick is asymmetric quantisation. Only the routed expert tensors are compressed to 2 bits, while attention layers, shared experts and the output head stay at higher precision. The design keeps the model’s most important computation paths precise.

On a 128 GB Mac with the q2 build, the engine has recorded prefill throughput of about 790 tokens per second and generation of about 39 tokens per second at a 2,048-token context, according to the project’s benchmarks.

The C-based engine supports Metal, CUDA and ROCm backends. It targets 64 GB Apple Silicon machines, Nvidia’s DGX Spark and AMD’s Strix Halo, with a context window of up to one million tokens.

DwarfStar 4 has spawned derivatives, including a Rust port for Apple Silicon. Its minimal, single-purpose design reflects Sanfilippo’s philosophy: compress the routed experts and keep the critical shared paths precise.