The q4 and q4f16 decoders keep only the MLP gate_proj and up_proj in int4 and every other MatMul, the LM head included, in int8, which brings them closer to the full-precision model. The vision encoders are unchanged. The card gets the new sizes and precision note.
