Unisound Releases U2, a Native Agentic Large Model for Complex Task Execution

Unisound has released U2, a general-purpose large language model designed for autonomous task execution, announced in a press release. The company describes U2 as a native agentic model capable of decomposing and completing more than 100 steps in complex workflows across areas such as office automation, software engineering, and research.

According to Unisound, U2 achieved high scores in several benchmark tests. It recorded 87.9 on GPQA Diamond for reasoning capability, 75 on SWE-Bench Verified for software engineering, 76.9 on Claw-Eval for autonomous agent execution, and 72.9 on GDPval for office productivity. These results place it among the top performers in its category.

U2 introduces a Hybrid Thinking mechanism that alternates between implicit and explicit reasoning depending on task complexity. The model also features Bounded Latent Rollout and Entropy-aware Switching, which allow dynamic adjustment of reasoning modes to maintain efficiency and control. It uses high-knowledge-density data screening and a knowledge distillation architecture to reduce redundancy while improving reliability.

The model is trained using an Agent-Harness collaborative framework, integrating feedback from real task trajectories to strengthen planning, tool use, and result verification. U2 focuses on three main capabilities: reasoning for logical stability, coding for engineering delivery, and agent execution for multi-tool collaboration. It is now available on the Unisound Token Hub for individuals, developers, and organizations.