5CD-AI/Vintern-1B-v3_5
Partial fine-tune vs OpenGVLab/InternVL2_5-1B. Partial change: 86% of the weights changed (attention, embeddings, lm_head, mlp, norms, other), in 24 of 24 layers.
Tensors changed: lm_head 1/1, Embeddings 4/5, Norms 96/145, MLP 174/174, Attention 192/264, other 48/48
7 derivatives checked
- rootonchair/Vintern-1B-v3_5-GGUF-ext: Same weights, converted (fewer layers)
- ngxson/Vintern-1B-v3_5-GGUF: Quantization (fewer layers)
- duyleekun/Vintern-1B-v3_5-BF16-GGUF: Same weights, converted (fewer layers)
- manhtran2323/vintern-handwriting-ocr: Adapter
- NPTIsMyName/vintern_1b_v3_5_finetuned: Adapter
- 5CD-AI/Vintern-Embedding-1B: LoRA-style
- dekthedev/Vintern-1B-v3_5-GGUF: Quantization (fewer layers)
3 other repos hold byte-for-byte identical weight files.