nvidia/Cosmos-Reason2-32B
Full fine-tune vs Qwen/Qwen3-VL-32B-Instruct. Fine-tune: 100% of the weights changed, across 64 of 64 layers.
Tensors changed: Embeddings 3/4, Norms 335/373, MLP 310/316, Attention 364/364, lm_head 1/1
4 derivatives checked
- mradermacher/Cosmos-Reason2-32B-i1-GGUF: Quantization (fewer layers)
- Hob-forge/Cosmos-Reason2-32B-GGUF: Quantization (fewer layers)
- mradermacher/Cosmos-Reason2-32B-GGUF: Quantization (fewer layers)
- DevQuasar/nvidia.Cosmos-Reason2-32B-GGUF: Quantization (fewer layers)