Qwen/Qwen2.5-Coder-3B-Instruct-GGUF
Same weights, converted vs Qwen/Qwen2.5-Coder-3B-Instruct. Same weights stored in another dtype (BF16 → F16/F32), plus 1 new tensors. Its layers carry AWQ-style scales (each input channel rescaled, with the norm before it compensating), which llama.cpp can fold in before quantizing: the same model.
Tensors changed: Embeddings 0/1, Norms 0/73, MLP 0/108, Attention 0/252