pentacoxian/qwen-qsa-llama.cpp is a llama.cpp branch that applies the Qwen Sparse Attention path to Qwen3.8-27B. Its README says it is configured to handle 262K context and image input on a single RTX 4090. The repository was created on November 4, 2026, and FACTS records no latest release. So this is a look at the implementation the repository presents, not at the changes of any particular release.
According to the README, the implementation includes memory savings, kernel changes, and DFlash2 draft rejection sampling. The repository also publishes speed and recall test results compared under the same weights and question conditions. It is written in C++ and uses the MIT license. The branch aims to handle long context and image input on a single GPU, and its key point is that it addresses the memory and kernel paths needed for that within llama.cpp-based code. That said, the descriptions and results in the README are what the repository author presents.
The results are hard to apply as-is to other devices or workloads. The README states that the recall evaluation is a simple check with 40 questions, and that no graphics card other than the RTX 4090 was tested. The published speed and recall results are likewise limited to comparisons under the same conditions, so treating them as evidence of a broad performance advantage would be premature. The current material can serve as a reference for the implementation direction, but the performance claims should be read conservatively until they are confirmed across more varied hardware and evaluations.