UniSound Joins Top Tier of Chinese LLMs with Token-Efficient U2 Foundation Model

UniSound (UniSound AI), the Hong Kong-listed AI company known for its speech recognition and AI capabilities, has entered the first tier of China's general-purpose large language model race with the launch of U2, a new foundation model that prioritizes token efficiency over raw parameter counts. U2 represents a deliberate departure from the industry's prevailing trend of scaling parameters and computing power. As the cost of inference and agent-based AI workloads escalates — driven by the token-intensive nature of chain-of-thought reasoning — UniSound has taken a different approach: maximizing intelligence per token rather than total model size. The strategy appears to be working. U2 achieves competitive or superior performance on key benchmarks — including IFBench for instruction following, Claw series Agent evaluations, and GPQA for hard reasoning tasks — while using significantly fewer activated parameters than comparable models. UniSound reports that U2 reduces thinking token consumption by approximately 25%, translating to substantially lower inference costs. The model's architecture is optimized for the agent era, supporting seamless integration with mainstream AI scaffolding frameworks including OpenClaw and Hermes. Independent tests show U2 completing complex tasks in fewer interaction rounds than competing models, with reduced error recovery cycles. In a practical demonstration, U2 generated a nearly 1,000-line interactive particle universe simulation as a single HTML file with no external dependencies — a task that typically separates strong models from weaker ones. The model also built a multi-module zodiac sign personality analysis application from vague user descriptions, demonstrating its ability to handle open-ended development tasks. UniSound positions U2 as a "general-purpose but efficiency-first" model — designed not just for benchmark performance but for real-world cost-effectiveness in production deployments. The model is now available through UniSound's Token Hub platform, supporting individual developers, enterprise teams, and organizations. The launch marks UniSound's strategic pivot toward becoming a foundational AI model provider, leveraging its decade-long expertise in speech and audio AI — accumulated since its founding in 2012 — to differentiate in an increasingly crowded Chinese LLM market. As the industry grapples with the economic challenges of scaling inference, U2's efficiency-first approach may point toward a more sustainable trajectory for large model deployment.