deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8

Full fine-tune vs deepseek-ai/DeepSeek-V3-Base. Fine-tune: 100% of the weights changed, across 61 of 61 layers.

Tensors changed: Attention 305/305, Norms 218/245, MLP 1050/1108

1 other repos hold byte-for-byte identical weight files.

On the Hub · Downloads on Model Pulse