Devotd Github
NewInference & serving

ingeniousidiocy/ds4-v41-m3ultra — DeepSeek V4.1 Flash on Metal

A DeepSeek V4.1 Flash repository built on antirez/ds4's DwarfStar, with Apple Metal optimizations and run configuration.

Repositoryingeniousidiocy/ds4-v41-m3ultra ↗

A DeepSeek V4.1 Flash repository built on antirez/ds4's DwarfStar, with Apple Metal optimizations and run configuration.

GitHub description: DeepSeek V4.1 Flash on the Apple M3 Ultra: Metal decode and prefill kernels, DSpark speculative decoding, KV disk store. Built on antirez/ds4 (upstream base bd66c40).

stars
★ 10
Forks
3
Language
C
License
MIT
Latest release
v41-video
Read original source ↗View on GitHub ↗

ingeniousidiocy/ds4-v41-m3ultra is a C repository optimized to run DeepSeek V4.1 Flash on Apple Metal. It is based on DwarfStar from antirez/ds4 and adds Metal decode and prefill kernels, DSpark speculative decoding, and the ability to keep KV data on disk. The upstream code it builds on is at commit bd66c40. The latest release tag reads v41-video, but the supplied material contains no changelog for that release.

The focus of this repository is to optimize the model execution and image input paths and to make each optimization comparable. The README explains that it groups the optimizations into 58 bundles and provides a switch to turn each one off. It also points to a run mode that prioritizes correctness, a reproduction procedure, and a document explaining the Metal implementation. According to the README, on a 512GB M3 Ultra the processing speed at an 8K context is 41.4 tokens per second. The license is MIT.

That processing speed, however, is the repository's own measurement as recorded in the README, not an independently verified figure. The value presented is limited to the condition of a 512GB M3 Ultra with an 8K context. The README states that no measurements for other hardware conditions exist, and that later changes may make the figure in the comparison image differ from the current default settings. The result is therefore a reference value for that configuration, not a standard that guarantees performance in other environments.