Antirez’s DwarfStar runs DeepSeek V4.1 Flash locally on 128 GB Macs
Salvatore Sanfilippo, the creator of Redis, is developing DwarfStar, a specialised inference engine for running selected large language models on personal hardware. Its documentation says DeepSeek V4.1 Flash can run with SSD streaming on one 128 GB Mac or NVIDIA DGX Spark; a community-made quantised build uses an importance matrix calibrated on the 1017 checkpoint. DwarfStar is beta software, and V4.1 support has limits across hardware and vision.
Bottom line — DwarfStar’s V4.1 guide lists SSD streaming on one 128 GB machine, with fast local storage required for the model’s Engram tables.
Go deeper 6
-
The Hugging Face build is a 3.91-bit-per-weight quantisation, with routed experts at 2.25 bits and most other components at 8 bits or more.
-
Hugging Face says the model file is about 366 GB and includes 189 GiB of Engram tables, which are read from SSD on demand rather than loaded into memory.
-
The builder says 32 of the model’s 15,360 routed experts received no tokens during calibration and were quantised without importance weighting.
-
DwarfStar’s guide says V4.1 vision works on Metal, but CUDA vision is not yet supported; the matching vision encoder is required.
-
The guide says V4.1 supports adjustable reasoning levels from 1 to 100; setting the level to zero disables reasoning.
-
The GitHub project describes itself as beta quality and warns that regressions and instability are possible.