Antirez demos DeepSeek V4.1 Flash at 2-bit on DwarfStar

Salvatore Sanfilippo, the Italian developer better known as antirez and the original author of Redis, has posted a new Italian-subtitled walkthrough of DeepSeek V4.1 Flash running at 2-bit under his own inference engine, DwarfStar (ds4). The engine is written in C with Metal, CUDA and HIP/ROCm backends and, per its own documentation, "not a generic GGUF runner": it targets a narrow set of DeepSeek and GLM models on purpose.

The trick that lets a Flash-class mixture-of-experts model fit on a single consumer machine is an asymmetric quantization scheme. The DwarfStar site sums the design up in one line: "Compress the routed experts, keep critical shared paths precise." In practice only the routed expert tensors are 2-bit; attention, shared experts and the output head stay at Q8_0, and the compressor/indexer path stays at F16. On the published M5 Max reference with 128 GB of unified memory, the same page records "prefill throughput of 790.2 tokens/second and generation at 39.4 tokens/second" for the q2 build at a 2,048-token context.

Two of the researchers we follow were passing the video around within a day of its posting, part of a wider run of interest as ds4 picks up ways to run near-frontier open weights on a single high-memory Mac or workstation GPU.