Description
Multimodal models answering about images guess at unclear details and never look things up. DeepEyesV2 is an agentic multimodal model that thinks with images, cropping and zooming, running code and searching the web while reasoning.
Trained with cold start and RL, with checkpoints available, it excels at visual tasks needing close looks and verification.
Thinking with images:Crop and zoom.
Tools:Code and search.
RL:Learns when to use tools.
Trained with cold start and RL, with checkpoints available, it excels at visual tasks needing close looks and verification.
Features
Thinking with images:Crop and zoom.
Tools:Code and search.
RL:Learns when to use tools.
