Unisound Launches U2-Flash Model; Shares Rise as Much as 10%

Unisound (09678) has launched U2-Flash, a new-generation high-performance flagship model. The company announced the official release of the new high-density intelligence model, which was developed through reinforced post-training based on the U2 general base model. Compared with the previous-generation U2 model, U2-Flash shows significant improvement in task completion quality and overall efficiency. On the DeepSWE v1.1 benchmark, which represents coding capability, U2-Flash doubled the score of its predecessor, reaching 64.6 points and surpassing models including GLM5.3-Flash and DeepSeek-V4-Pro-0813.

The model achieves comprehensive optimization in end-to-end execution efficiency and inference cost. Agent task iteration steps are reduced by 20% to 30%, task execution cycles are shortened by 35%, and token consumption is reduced by 20% to 30%. U2-Flash adopts a sparse mixture-of-experts (MoE) architecture, with total parameters of approximately 266 billion and about 10 billion parameters activated per inference. Average time to first token is kept within 3 seconds, and peak output throughput reaches up to 300 tokens per second. The model unifies capabilities across coding, agents, mathematical reasoning, and instruction following within a single set of weights, achieving flagship-level task completion quality with an activation scale far smaller than comparable dense models.

Unisound shares rose as much as 10%, reaching a high of HK$77.25. As of 11:07 a. m. , the stock was at HK$72.25, up 3.1%, with turnover exceeding RMB 260 million.