Unisound AI Launches U2 LLM: 266B Parameters, Agent-Driven

Unisound AI Technology has officially released U2, a new large language model. This model uses a Mixture of Experts architecture and has 266 billion parameters. It aims to boost the effective knowledge per token and the value from business applications. U2 is also an Agent-driven model. It can handle full-stack development, intelligent orchestration, and deep inference. It uses a native inference path distillation technology to perform task decomposition, execution, validation, and optimization. This makes it a strong foundation for new AI Agents. The model features a hybrid reasoning mode and a synchronous training mechanism. This reduces token consumption per task and improves overall token utilization. U2 shows strong performance in areas like instruction following, long-context processing, and scientific reasoning, comparable to other leading models. This release marks a new step in the company's AI commercialization.