tencent/Penguin-VL-2B
Full fine-tune (fewer layers) vs Qwen/Qwen3-1.7B. Qwen/Qwen3-1.7B matches part of tencent/Penguin-VL-2B (310 of its 625 tensors; the rest of tencent/Penguin-VL-2B has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 28 of 28 layers.
Tensors changed: Embeddings 1/1, Norms 108/113, MLP 84/84, Attention 112/112
1 other repos hold byte-for-byte identical weight files.