0xSero/DeepSeek-V4-Flash-0731-REAP
Partial fine-tune (fewer layers) vs deepseek-ai/DeepSeek-V4-Flash-0731. 0xSero/DeepSeek-V4-Flash-0731-REAP matches part of deepseek-ai/DeepSeek-V4-Flash-0731 (45732 of its 72317 tensors; the rest of deepseek-ai/DeepSeek-V4-Flash-0731 has no counterpart). On the shared part: Partial change: 17% of the weights changed (mlp), in 43 of 43 layers.
Tensors changed: Embeddings 0/1, MLP 840/1259, Attention 0/893, Norms 0/249, other 0/12