freriadoc/avalon is a llama.cpp fork tailored to Qwen3.8-Flash-Next. The repository description and README cover a configuration targeting this model. No registered latest release appears in FACTS, so there is nothing to review as version-by-version updates. The repository is written in C++ and uses the MIT license. The focus of this project is an execution path tuned to a specific model rather than broad support for many models.
According to the README, Avalon uses ROCm on AMD GPUs and describes an approach that keeps some of the expert weights in system memory. The model activates 10 of 512 experts, and mixture of experts (MoE) means only some experts take part in the computation depending on the input. The target configuration Avalon presents is to split those weights so the model can be handled on a single consumer GPU. This is a case of a single-model optimization fork addressing hardware memory constraints.
The README presents conditional generation speeds on an RX 7900 XT and long-context experiments. Comparison results with Strata and the RTX 4090 appear in a separate section, but FACTS contains no specific measurements. The README notes that several of the numbers are results of the author's environment and measurement procedure, so they cannot be read as a universal performance guarantee. The verifiable scope is therefore limited to the specific model and the stated test conditions, and there is no basis for assuming the same results on other GPUs or models.