text_proj_embed, talker_codec_embedding and code_predictor_codec_embeddings were raw float32: 2.28 GB of a 4.91 GB bundle for tables that are read row-wise and copied straight into the talker prefill. Stored as float16 they are 1.14 GB. Round-trip error measured on the shipped tensors: max 3.3e-4 / mean relative 1.8e-4 on the text projection, and bit-exact on the two codec tables (their values were already float16-representable). Seven synthesis comparisons totalling ~46 s of audio, sampled and greedy, came out byte-identical to the float32 run. The runtime picks the dtype from the file size, so old and new files both load.
