Avalon for Windows (CUDA 13.4, x64) - windows-v1

Native Windows build of Avalon for Qwen3.8-Flash-Next on NVIDIA GPUs. Experts stay in system RAM and the rest runs on the GPU, so the full model runs on one 16 GB card.

Source: branch win-release (Avalon main 12abd35 plus three Windows fixes, offered upstream as pull requests: the MSVC build, CUDA_MODULE_LOADING=EAGER by default, and a CUDA-graphs hang fix).

CPU code: AVX2. GPU code: RTX 50 (sm_120a), RTX 40 (sm_89), RTX 30 (sm_86). Tested on an RTX 5070 Ti 16 GB only.

Needs an NVIDIA driver that supports CUDA 13.4 (tested 617.42), the Microsoft Visual C++ Redistributable (x64), and more system RAM than the model file.

Includes ready launchers for six builds: Avalon IQ3_KS, Unsloth UD-IQ4_XS, UD-Q6_K_XL and Q8_0, Swift-1.5 Q6_K and Q8_0. Model files are not included; README.txt has the download commands.

On our machine (RTX 5070 Ti 16 GB, Ryzen 9 9950X, 192 GB DDR5), Avalon wrote about twice as fast as stock llama.cpp on the same files and read long prompts 7.5-12x faster. A 53K-token document's first word arrived in 22-33 seconds instead of 3-6.5 minutes. Full tables are in README.txt.