XPENG-AI/X-AuT
Partial fine-tune vs Qwen/Qwen3-ASR-0.6B. Partial change: 71% of the weights changed (attention, embeddings, lm_head, mlp, norms, other), in 28 of 28 layers.
Tensors changed: other 11/11, MLP 56/140, Norms 57/171, Attention 224/224, lm_head 1/1, Embeddings 1/1