Baidu ERNIE 4.5 VL MoE vision-language models (28B-A3B, 424B-A47B) with heterogeneous text/vision experts