Jwuthrich/selfjev-4b
Adapter vs Qwen/Qwen3.5-4B. Jwuthrich/selfjev-4b is a LoRA adapter, not full weights: rank 64, on in_proj_a, in_proj_b, in_proj_qkv, in_proj_z, k_proj, o_proj, out_proj, q_proj… in 32 layers (152 adapted matrices). It modifies Qwen/Qwen3.5-4B at load time.
1 other repos hold byte-for-byte identical weight files.