MK4-Research/MK4-Ginko-v2
Partial fine-tune vs mlx-community/Qwen3.8-27B-4bit. Partial change: 70% of the weights changed (attention, embeddings, lm_head, mlp), in 64 of 64 layers.
Tensors changed: Embeddings 1/2, Norms 0/209, Attention 457/752, MLP 266/384, lm_head 2/2