tencent/Penguin-VL-8B
Full fine-tune (fewer layers) vs Qwen/Qwen3-8B. Qwen/Qwen3-8B matches part of tencent/Penguin-VL-8B (399 of its 714 tensors; the rest of tencent/Penguin-VL-8B has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 36 of 36 layers.
Tensors changed: Embeddings 1/1, Norms 144/145, MLP 108/108, Attention 144/144, lm_head 1/1