ibm-granite/granite-vision-3.2-2b
Full fine-tune (fewer layers) vs ibm-granite/granite-3.1-2b-instruct. ibm-granite/granite-3.1-2b-instruct matches part of ibm-granite/granite-vision-3.2-2b (361 of its 815 tensors; the rest of ibm-granite/granite-vision-3.2-2b has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 40 of 40 layers.
Tensors changed: Norms 81/81, MLP 120/120, Attention 160/160
6 derivatives checked
- bartowski/ibm-granite_granite-vision-3.2-2b-GGUF: Quantization (fewer layers)
- ibm-research/granite-vision-3.2-2b-GGUF: Same weights, converted (fewer layers)
- lmstudio-community/granite-vision-3.2-2b-GGUF: Quantization (fewer layers)
- DevQuasar-5/ibm-granite.granite-vision-3.2-2b-GGUF: Same weights, converted (fewer layers)
- unsloth/granite-vision-3.2-2b: Same weights, repacked
- unsloth/granite-vision-3.2-2b-unsloth-bnb-4bit: Quantization
1 other repos hold byte-for-byte identical weight files.