webbrain-one/webbrain-compass-tiny-xs-v3
LoRA-style vs XHToken/Spark-X2.5-1.7B. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, g_proj, gate_proj, out_proj, q_k_v_proj, up_proj changed, in 28 of 28 layers; norms, biases and embeddings are untouched.
Tensors changed: Embeddings 0/1, Norms 0/57, MLP 84/84, Attention 84/84