v0-1.5-md

v0-1.5-md is the medium-sized model in Vercel's v0 1.5 composite family, purpose-built for UI and front-end code generation. Unlike a single-purpose language model, it uses a three-layer architecture that splits work across retrieval, generation, and correction stages. The pipeline starts with retrieval-augmented generation pulling framework documentation and UI examples, then moves to Anthropic Sonnet 4 for the main generation pass, and finishes with Vercel's custom AutoFix model running a final correction sweep. This division of labor tackles failure modes differently than a standalone model would, addressing quality issues at each stage rather than relying on one model to handle everything.

The model achieved a 93.87% error-free rate on Vercel's web development evaluation—a figure that outpaced Claude 4 Opus at 78.43%, GPT-4.1 at 58.82%, and Gemini 2.5 Pro at 58.82% on the same test. Interestingly, the medium-sized variant outperforms the larger v0-1.5-lg model, which reached 89.80% error-free output. Vercel attributes this to better alignment between the medium base model's capacity and the AutoFix layer's correction range, suggesting that scaling up base model size does not automatically translate to better end results for this particular task. The AutoFix model alone reaches 86.14% error-free output, showing how the correction stage contributes meaningful quality gains to the composite system.