Unisound AI Technology Co., Ltd. (Unisound) has officially released “U2”, a proprietary large language model built on a Mixture-of-Experts (MoE) architecture and equipped with 266.00 billion parameters. Announced on 8 June 2026, U2 integrates fast- and slow-thinking paradigms to raise the “effective knowledge payload per Token” and lower inference costs, forming the company’s new technological core for commercial AI deployment.

U2 is engineered as a natively agent-driven model, capable of autonomous task decomposition, execution, validation and optimisation through native inference path distillation. An innovative hybrid-reasoning mode and a model-Harness synchronous-training mechanism further cut Token consumption per task and increase overall utilisation efficiency, positioning U2 for broad sector integration.

Authoritative professional evaluations rank U2 on par with global leading models across key benchmarks, including instruction following, long-context processing, scientific reasoning, software-engineering agents and specialised toolchain adaptation (Claw evaluation).

Strategically, Unisound plans to leverage U2’s capabilities to: 1) scale existing verticals such as healthcare, insurance and transportation through rapid business expansion and model upgrades; and 2) deepen participation in standardised cloud-based services via large-model API offerings, opening new incremental market opportunities.

The official launch marks an upgraded commercialisation pathway for Unisound, reinforcing its objective to unlock AI value across diverse application scenarios.