Shoestring (@Shoestring_Dev) header
Shoestring (@Shoestring_Dev) profile picture

back in the lab again, yo

Joined October 2026
Photo by Shoestring (@Shoestring_Dev)Photo by Shoestring (@Shoestring_Dev)Photo by Shoestring (@Shoestring_Dev)
Two weeks ago, two Strix Halo mini PCs ran the full DeepSeek V4.1 Flash at 52.8 tok/s. Since then StrixFold learned to: - see images - read 1M tokens - keep its cache on disk through restarts - run on a Bosgame M5 pair at 77-80% of two DGX Sparks Free and open source 👇
1
33
A lot of it came from the community: RDMA transport, image input, the disk cache and a shared memory pool are all outside PRs. Thank you TheBeaninator, dturini12, cc-bjojac and everyone on the Discord who sent tests and numbers.
13
StrixFold just passed 35 stars on GitHub. Thank you. New today: drafting is on by default (~1.7x faster, same output), and we measured speed at long context. It holds flat from 0 to 48K, plus a 1M-token run. DeepSeek V4.1 Flash on 2xStrix Halo over USB4. github.com/Shoestring-dev/St…
39
StrixFold is now open source. It runs DeepSeek-V4.1-Flash on two AMD Strix Halo boxes linked over USB4. 53 tok/s for one chat, 1,543 tok/s prompt reading, 108 tok/s across eight chats. 72% of two DGX Sparks on the same weights. @ashxhart github.com/Shoestring-dev/St…
2
31
@turboderp_ we are using your weights for it
23
Building. Building. Building.
18