bitexploder/fnfl-v100 is a C++ inference engine built to run Qwen Flash Next 3.8 on Tesla V100 GPUs. It is derived from llama.cpp, and the repository description states that the code was written from scratch for the V100. It targets both single-GPU and multi-GPU configurations. There is no information about the latest release.
This engine focuses on a specific model-and-GPU combination rather than being a general-purpose inference tool. According to the repository guide, it supports two quantization formats and provides mixture-of-experts caching and concurrent request handling. A mixture of experts is a structure where only some of several expert models are selected to do the processing, and caching is a way of storing the needed experts so they can be reused quickly. It also supports image input, and includes the ability to read long prompts while other users' generations are in progress. The license is MIT. The README contains build and run instructions, version-by-version changes, and a performance comparison table that states the hardware conditions.
The performance comparison table shows measurements published by the repository and is not a result that has been independently reproduced and verified. It is therefore more appropriate to treat the table as reference material under the conditions stated by that repository. The verifiable scope also extends only to the feature notes specific to Qwen Flash Next 3.8 and the V100. No evidence is presented that the same results can be expected on other models or hardware.