inferencerlabs/Bonsai-8B-MLX-Q2
Partial fine-tune vs prism-ml/Bonsai-8B-mlx-1bit. Partial change: 50% of the weights changed (attention, embeddings, lm_head, mlp), in 36 of 36 layers.
Tensors changed: lm_head 1/2, Embeddings 1/2, Norms 0/145, MLP 108/216, Attention 144/288