deepseek-ai/DeepSeek-V3.2-Speciale

Full fine-tune vs deepseek-ai/DeepSeek-V3.2-Exp-Base. Fine-tune: 100% of the weights changed, across 62 of 62 layers.

Tensors changed: Embeddings 2/2, Attention 790/930, Norms 376/376, MLP 1554/1689, lm_head 1/1, other 2/2

On the Hub · Downloads on Model Pulse