OpenGVLab/InternVL3-38B
Full fine-tune vs OpenGVLab/InternVL3-38B-Instruct. Fine-tune: 99% of the weights changed, across 64 of 64 layers.
Tensors changed: Embeddings 4/5, Norms 231/309, Attention 581/583, other 87/90, MLP 370/378, lm_head 1/1
3 derivatives checked
- unsloth/InternVL3-38B-GGUF: Same weights, converted (fewer layers)
- dmis-lab/InternVL3-38B-CVRR: Different architecture
- Skywork/Skywork-R1V3-38B: Full fine-tune
2 other repos hold byte-for-byte identical weight files.