More information and benchmarks will be posted soon.
This was converted from the original Qwen/Qwen3.8-27B repository.
It was quantized using my
advanced-gguf-quantizer tool.
To use this model, you must build my unofficial MXFP8 fork of llama.cpp at:
https://github.com/michaelw9999/llama.cpp/tree/mxfp8_cuda
On Blackwell, I am seeing that MXFP8 is faster than Q8_0 and is comparable in quality.
I will post some mixed NVFP4/MXFP6/MXFP8 models soon that should provide the fastest possible Blackwell models.
Please let me know your feedback and if you have any problems, I will be happy to help!
