openinterx/UGC-VideoCaptioner
LoRA-style vs Qwen/Qwen2.5-Omni-3B. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, gate_proj, k_proj, o_proj, q_proj, up_proj, v_proj changed, in 20 of 36 layers; norms, biases and embeddings are untouched.
Tensors changed: other 0/549, MLP 108/504, Norms 0/363, Attention 144/1076, Embeddings 0/51, lm_head 0/1
2 other repos hold byte-for-byte identical weight files.