thinkingmachines/Inkling-NVFP4
Partial fine-tune (fewer layers) vs thinkingmachines/Inkling. thinkingmachines/Inkling matches part of thinkingmachines/Inkling-NVFP4 (1426 of its 2056 tensors; the rest of thinkingmachines/Inkling-NVFP4 has no counterpart). On the shared part: Partial change: 0% of the weights changed (mlp), in 64 of 66 layers.
Tensors changed: MLP 128/426, Norms 0/310, Attention 0/666, other 0/13, Embeddings 0/11
2 derivatives checked
- mlx-community/Inkling-NVFP4-mlx-4bit: Quantization
- modal-labs/Inkling-NVFP4-DFlash: Different architecture