RAS-LAB/Qwen3-Coder-16B-A3B-KREAP-MLX
LoRA-style (fewer layers) vs Qwen/Qwen3-Coder-30B-A3B-Instruct. RAS-LAB/Qwen3-Coder-16B-A3B-KREAP-MLX matches part of Qwen/Qwen3-Coder-30B-A3B-Instruct (9603 of its 18867 tensors; the rest of Qwen/Qwen3-Coder-30B-A3B-Instruct has no counterpart). On the shared part: Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, gate_proj, up_proj changed, in 48 of 48 layers; norms, biases and embeddings are untouched.
Tensors changed: Embeddings 0/1, Norms 0/193, MLP 1284/1310, Attention 0/192, lm_head 0/1