slimsami/strycast is a community fork of the Gufo inference engine for AMD Strix Halo. It covers model setups that are not in Gufo's model list, matched to smaller memory environments or specific use cases. The latest release is v0.4.0-qwen35moe.1. The repository description covers Qwen3.6-35B-A3B with MTP and DFlash2, and the release tag also points to a Qwen 35B MoE setup. However, this record contains no detailed list of the changes added in that release.
The scope of this project is to broaden the model setups that can be chosen on Strix Halo hardware and the guidance for running them. The README links the Qwen3.6-35B-A3B and Ornith 1.5 122B guides to inference methods, weight files, performance tables, and quality materials. The repository description states that it supports up to 3,000 tokens per second in the prefill stage for Qwen 35B-A3B. Prefill is the stage that processes the input content first. This is the performance information presented by the repository, and it does not describe the performance of other models or hardware.
The implementation language is C++ and the license is MIT. The README also explains the contribution rule that only changes which pass testing are merged into the main branch. However, the guides and the performance and quality materials presented are focused on the setups of these two models. So Strycast is closer to a fork worth consulting when looking into specific model setups that Gufo does not cover, rather than an engine that encompasses every Strix Halo inference configuration.