badtheorylabs/Tinfield-1

Partial fine-tune vs Qwen/Qwen3.8-Flash-Next. Partial change: 2% of the weights changed (attention, mlp, other), in 48 of 48 layers.

Tensors changed: Norms 0/302, other 2/8, Attention 372/608, MLP 144/602, Embeddings 0/137, lm_head 0/1

4 derivatives checked