A source comparison and a failed parity run found that LeRobot 0.6.1 with transformers 5.3.0 scales pi0 image embeddings about 45x too small. 0.6.1 needs transformers 5.4 or newer. With 5.5.4 and a ported mask patch, our pi0 export matched the old stack (cosine 0.9999997). If you run pi0 or pi0.5 inference, which transformers version is in your environment, and do your action outputs look right? I can share the comparison script.
DeepSeek releases updated V4.1-Flash model it says completes tasks faster