tencent/UI-Mate-democua-27B

Full fine-tune vs Qwen/Qwen3.6-27B. Fine-tune: 95% of the weights changed, across 64 of 64 layers.

Tensors changed: Embeddings 3/4, MLP 285/304, Attention 492/556, lm_head 1/1, Norms 221/319

3 derivatives checked