Which local models actually work with Cline? AMD tested them all

AMD's guide to local "vibe coding" with Cline, LM Studio, and VS Code provides insights into the models and hardware configurations suitable for coding tasks. After testing over 20 models, they found that only a few reliably work, with smaller models often producing broken outputs. The guide details how RAM and VRAM affect model performance, emphasizing that system RAM is crucial for loading models, while VRAM influences inference speed. It recommends the GGUF format for Windows, Linux, and Mac users for broader compatibility, while MLX is suggested for those exclusively using Mac with Apple Silicon. Quantization is discussed as a method to save memory by reducing model precision, with 4-bit quantization deemed sufficient for production-ready coding tasks. The guide outlines RAM requirements for different models, with 32GB as the minimum viable tier and 128GB+ offering cloud-level performance. Platform-specific configurations are provided for Windows, Mac, and Linux users, and AMD highlights that models smaller than Qwen3 Coder 30B are unsuitable for Cline due to their inability to handle autonomous coding tasks effectively.