valeriodolci/capy-flash is a project that runs Qwen3.8-Flash-Next (MoE, NVFP4) and Qwen3.8-27B on two 16GB consumer GPUs. The v0.6.0 release published on GitHub improves graphics memory usage and expert module placement in long conversations. It also adds short input handling and optional image understanding. These are the changes in the new version of the repository.
This project describes a tensor parallel configuration that uses both GPUs together and a method of offloading expert modules to the CPU. The models listed in the repository description are Qwen3.8-Flash-Next (MoE, NVFP4) and Qwen3.8-27B, and the project was previously named ninfer-tp2. The summary of this release lists improved graphics memory usage and expert module placement for long conversations as the core changes, and presents short input handling and optional image understanding as added features. Implementation details and the supported scope of each feature are not given in the available information.
Speed comparisons for the new processing method have also been published, but the figures are the project's own test results and are not independently verified. So they cannot be taken as evidence that the same speed will be achieved in other environments. The configuration presented in the repository description is two 16GB consumer GPUs, and results on other GPU configurations are not in the available information. The significance of this change is that long conversation handling, expert placement, and short input plus optional image understanding features have been added, and the speed comparison should be read as a project-side result only until it is independently verified.