$PHA
Official FP8 Weight: 755 GB
Phala W4AFP8 Weight: 368 GB
This frees up 387 GB for KV caching, CUDA charts, runtime buffers, and service margin under a full 1M token window.
Official FP8 Weight: 755 GB
Phala W4AFP8 Weight: 368 GB
This frees up 387 GB for KV caching, CUDA charts, runtime buffers, and service margin under a full 1M token window.

