nvidia/Cosmos-Reason1-7B
Full fine-tune vs Qwen/Qwen2.5-VL-7B-Instruct. Fine-tune: 93% of the weights changed, across 32 of 32 layers.
Tensors changed: Embeddings 2/2, Norms 63/122, MLP 280/280, Attention 324/324, lm_head 0/1
6 derivatives checked
- unsloth/Cosmos-Reason1-7B-GGUF: Same weights, converted (fewer layers)
- mradermacher/Cosmos-Reason1-7B-i1-GGUF: Full fine-tune (fewer layers)
- mradermacher/Cosmos-Reason1-7B-GGUF: Full fine-tune (fewer layers)
- DevQuasar/nvidia.Cosmos-Reason1-7B-GGUF: Full fine-tune (fewer layers)
- unsloth/Cosmos-Reason1-7B: Identical copy
- sasa2000/Cosmos-Reason1-7B-heretic: Abliterated
1 other repos hold byte-for-byte identical weight files.