CFENG + MLAThe shape
The shape
of a model.
Character model · 65-symbol vocabulary · D = 128
ARCHITECTURE / LIVE PARAMETER ACCOUNTING
CHARACTERS65 → 128
↓MULTISCALE CARRY10 past-only streams · FiLM
↓3 × [ MLA + CFENG ]width 280 · 5 rational levels
↓NORMALIZE → PREDICT280 → 65
5,307,665trainable parameters
Approximately 81 MiB for parameters, gradients, and two FP32 Adam states. Activations, optimizer variants, and runtime overhead are excluded. This tool calculates structure; it does not train a model.