Summary & Key Takeaways ​

• Simon Willison conducted a direct comparison between a locally hosted Qwen3.6-35B-A3B model and Anthropic's Claude Opus 4.7 for image generation.
• Surprisingly, the local Qwen model produced a more impressive pelican image than the commercial Claude Opus 4.7.
• This highlights the growing capabilities of smaller, locally runnable AI models.

Our Commentary ​

This is exactly the kind of real-world, anecdotal comparison we love to see. It challenges the perception that only the largest, cloud-hosted models can deliver top-tier results. The fact that a model running on a laptop can outperform a commercial API for a specific task is a powerful statement about the progress in local AI and the potential of open-source models. It makes us wonder what other 'underdog' models are out there waiting to surprise us.