Windows (CUDA) build available: prebuilt zip and launchers

[newmilltechnology]

I've published a native Windows build of Avalon for NVIDIA GPUs: https://github.com/newmilltechnology/Avalon/releases/tag/windows-v1

Avalon (https://github.com/freriadoc/Avalon) is freriadoc's llama.cpp fork for Qwen3.8-Flash-Next. It keeps the model's experts in system RAM and the rest on the GPU, so the full model runs on one 16 GB card.

This build is Avalon main at 12abd35d7 plus three Windows fixes I've offered upstream as pull requests (MSVC build, CUDA_MODULE_LOADING=EAGER by default, and a hang fix with CUDA graphs). AVX2 CPU code; GPU code for RTX 50, 40 and 30 series. Tested on an RTX 5070 Ti 16 GB with 192 GB RAM only.

The zip includes launchers for Avalon's own IQ3_KS release and for five community quants (Unsloth UD-IQ4_XS, UD-Q6_K_XL, Q8_0; Swift-1.5 Q6_K, Q8_0), with the settings each one needs. On my machine IQ3_KS writes about 79-85 tok/s and reads about 2,300-2,500 tok/s; the 6-bit and 8-bit files run about 2x faster writing and 7.5-12x faster reading than stock llama.cpp on the same files. Full numbers are in the release's README.

I'm running Swift-1.5 Q8_0 on it as my primary coding model, and it's wonderfully fast.