crgregersen/m10dis
Qwen3.8-Flash-Next-FP8 (48-layer MoE, 512 experts) on one RTX 3070 Ti plus 24 Maxwell GPUs from six Tesla M10 cards. PCIe multicast, GPU page-table remapping, 96 tok/s decode.
GitHub repository with 38 stars and 5 forks.
Language: Cuda