cerebras/DeepSeek-V3.2-REAP-345B-A37B

Partial fine-tune (fewer layers) vs deepseek-ai/DeepSeek-V3.2. cerebras/DeepSeek-V3.2-REAP-345B-A37B matches part of deepseek-ai/DeepSeek-V3.2 (46194 of its 92425 tensors; the rest of deepseek-ai/DeepSeek-V3.2 has no counterpart). On the shared part: Partial change: 18% of the weights changed (mlp), in 58 of 61 layers.

Tensors changed: Embeddings 0/1, Attention 0/915, Norms 0/367, MLP 587/960, lm_head 0/1

On the Hub · Downloads on Model Pulse