ByteDance/Sa2VA-InternVL3-8B

Full fine-tune (fewer layers) vs OpenGVLab/InternVL3-8B. OpenGVLab/InternVL3-8B matches part of ByteDance/Sa2VA-InternVL3-8B (683 of its 1589 tensors; the rest of ByteDance/Sa2VA-InternVL3-8B has no counterpart). On the shared part: Fine-tune: 96% of the weights changed, across 28 of 28 layers.

Tensors changed: Norms 0/153, MLP 90/186, Attention 112/292, Embeddings 0/4, other 0/48

On the Hub · Downloads on Model Pulse