Well, I believe language models usually use 2-byte (16 bit) tokens, which corresponds to an embedding dimension of 2^16=65536. With 400 bytes per token this would be 2^(400*8), which is an extremely large number. Way too large to be practical, I assume.