donedynamics/Qwen3.8-27B-heretic-MLX-bf16
Partial fine-tune (fewer layers) vs trohrbaugh/Qwen3.8-27B-heretic-ara. donedynamics/Qwen3.8-27B-heretic-MLX-bf16 matches part of trohrbaugh/Qwen3.8-27B-heretic-ara (802 of its 1199 tensors; the rest of trohrbaugh/Qwen3.8-27B-heretic-ara has no counterpart). On the shared part: Partial change: 0% of the weights changed (norms), in 64 of 64 layers.
Tensors changed: Embeddings 0/1, Norms 161/209, Attention 0/400, MLP 0/192