Machine-Readable Model Cards Are Becoming the AI Discovery Layer

The old model card was built for humans The first wave of model cards did an important job: they gave AI providers a place to describe what a model was, how it was trained, where it performed well, and where it might fail. They were designed for researchers, journalists, policy teams, and enterprise buyers trying to understand a model beyond a leaderboard score. But the AI market has changed. There are now hundreds of general-purpose LLMs, specialist coding models, embedding models, image generators, rerankers, speech models, multimodal systems, and agent-oriented models. Buyers are not reading every paper or scrolling every launch page. Developers are not manually comparing context windows, function-calling formats, rate limits, latency regions, tool support, and data-retention policies every time they build a feature. The model card is becoming less like a brochure and more like infrastructure. The next version will not just be readable by humans. It will be readable by software. Why model discovery now needs structured data AI model discovery used to be simple enough to handle through blog posts, social media launches, and benchmark tables. That worked when the question wa