stepfun-ai/StepFun-Prover-Preview-32B

Full fine-tune vs deepseek-ai/DeepSeek-R1-Distill-Qwen-32B. Fine-tune: 98% of the weights changed, across 64 of 64 layers.

Tensors changed: Embeddings 0/1, Attention 448/448, MLP 192/192, Norms 129/129, lm_head 1/1

1 derivatives checked