Computer-use capability & efficiency
A model family for agentic productivity, with Pro and Turbo variants supporting real-world workflows, code engineering, and multimodal understanding.
Competitive programming results
A unified model family for complex real-world tasks, with reasoning, multimodal, and agent capabilities.
Reasoning performance & efficiency
A model release focused on generalized real-world agency.
Unified agent capabilities
An agent model combining GUI, game, code, and tool-use capabilities through multi-turn reinforcement learning.
GUI benchmark performance overview
A native agent model for automated interaction with graphical user interfaces.
* Equal contribution.
Visual-native search & data evolution
A visual-native search framework that preserves reusable image evidence and evolves training data through policy-rollout feedback.
Learning strategic tool use
Reinforcement learning for multi-turn tool use, enabling language models to learn when and how to call tools during reasoning.
Difficulty-aware reasoning
Adaptive and controllable reasoning through difficulty-aware allocation of reasoning budgets.
Knowledge-guided self-learning
A self-learning approach that makes the knowledge within language models explicit to support further learning.
Tool-use benchmark construction
A benchmark spanning tool planning, creation, and use in complex real-world scenarios.