On September 15, UniSound announced the launch of U2-Flash, a new high-density intelligent model using a sparse mixture-of-experts architecture with about 266 billion total parameters and roughly 10 billion activated per inference. On the DeepSWE v1.1 coding benchmark it scored 64.6, surpassing GLM5.3-Flash and DeepSeek-V4-Pro-0813; its TerminalBench 3.0 score rose to 24.3, ahead of trillion-parameter-class models.
The model has completed systematic adaptation to mainstream domestic compute platforms. Agent-task iteration steps fell 20–30% and Token consumption dropped 20–30%, pointing to faster commercialization of AI applications on domestic hardware. The news lifted Hong Kong's hard-tech stocks, with UniSound up over 9%.
Analysts at CSC maintain a long-term bullish view on the domestic AI-compute chain, noting North American hyperscalers' combined Q2 2026 capex of about $171.2 billion, up 78.6% year on year, and strong optical-module and GPU/ASIC earnings that confirm booming AI-compute demand.
Source: 21st Century Business Herald
