franzmoca/gufo is an inference engine that runs local models on Strix Halo hardware under Linux. This repository is a fork of Gufo extended to run Gemma 4, adding Gemma 4-specific compute kernels and multi-token prediction. The repository also states that image input and tool calling are supported. No public release version is given in FACTS.
The core of this project is not a general-purpose inference engine but a model execution path tailored to Strix Halo. Multi-token prediction is a feature that predicts several upcoming tokens at once, and the repository explains that it uses this to improve throughput. The per-model documentation links to performance measurements and quality reports. According to the figures recorded in the repository, the 8-user throughput of Qwen Flash Next Q4_K_XL is 157.22 tok/s, and Qwen27B Q4_K_XL reaches 70.56 tok/s under single-user conditions with DFlash2. The primary language is C++ and the license is MIT.
These figures are records for specific models and the Strix Halo environment, so they cannot be generalized to the performance of other hardware or models. Support is also limited to Strix Halo on Linux. The repository states that it periodically merges the upstream repository, so when reviewing this fork, the hardware limitation and its relationship to the upstream repository should be taken into account.