Add final model config, tokenizer and query tokens, chat template, image/video processors, and inference configuration. Existing checkpoint files remain unchanged. Verified offline loading and synthetic end-to-end inference with original Qwen weight loading forbidden.
Release final CARE-Drive backbone and coupled experts
Exact model-tensor exports of A0-DYNAMIC-L1, C3-XY-EP1 and D1-H-ADAPT epoch 5. Optimizer and internal run metadata omitted. All source hashes, tensor equality, and published-code strict loaders verified.