ERNIE-4.5-21B-A3B-Thinking

Baidu continues to release open models. This one is their first open reasoning model, supporting 128,000 tokens as context. Furthermore, they go a step towards the community, releasing weights directly usable in transformers (and thus, vLLM etc.). Previously, they split the model for both PyTorch and PaddlePaddle, which, while usable, added a lot of clutter. The benchmarks are also impressive for its size.