Odin runs all 32 Llama-3-8B layers under fully homomorphic encryption in 366 seconds.

The design co-designs CKKS ciphertext packing with model execution using a feature-major cross-layer layout, so QK^T scores feed softmax directly and PV consumes probabilities with no intermediate repacking. Nonlinear ops use minimax polynomial approximation with joint error allocation guided by model quality. 366.4s and 58.9 GiB peak on a single H100 80GB against THOR's 1651.9s, claimed as the first open-source end-to-end GPU CKKS implementation of Llama-3.