OpenGVLab/InternVL3-8B
Full fine-tune vs OpenGVLab/InternVL3-8B-Instruct. Fine-tune: 100% of the weights changed, across 28 of 28 layers.
Tensors changed: Embeddings 5/5, Norms 102/153, MLP 180/186, Attention 291/292, other 0/48, lm_head 1/1
15 derivatives checked
- unsloth/InternVL3-8B-GGUF: Same weights, converted (fewer layers)
- ByteDance/Sa2VA-Qwen2_5-VL-3B: Different architecture
- mradermacher/InternVL3-8B-i1-GGUF: Quantization (fewer layers)
- ByteDance/Sa2VA-Qwen3-VL-4B: Different architecture
- ByteDance/Sa2VA-InternVL3-2B: Different architecture
- mradermacher/InternVL3-8B-GGUF: Same weights, converted (fewer layers)
- ByteDance/Sa2VA-Qwen2_5-VL-7B: Different architecture
- InternSVG/InternSVG-8B: Different architecture
- ByteDance/Sa2VA-Qwen3-VL-2B: Different architecture
- sensenova/SenseNova-SI-1.1-InternVL3-8B: Partial fine-tune
- sensenova/SenseNova-SI-1.3-InternVL3-8B: Partial fine-tune
- ByteDance/Sa2VA-InternVL3-14B: Different architecture
- sensenova/SenseNova-SI-1.4-InternVL3-8B: Full fine-tune
- sensenova/SenseNova-SI-1.2-InternVL3-8B: Partial fine-tune
- ByteDance/Sa2VA-InternVL3-8B: Full fine-tune (fewer layers)
3 other repos hold byte-for-byte identical weight files.