a pretty fun and quick tests i do with vision models is to screenshot the hackernews homepage and ask the model to return a json representation of the screenshot - qwen 3.5 0.8b did surprisingly well at this.