UNISOUND (09678.HK): Launches High-Density Intelligent Model U2-Flash, Coding Benchmark Score Doubles
- The TerminalBench3.0 score rose to 24.3, while SWE-Bench Pro reached 61.6, up 10.5 points from the previous generation.
- The model uses a sparse mixture-of-experts architecture with about 266 billion total parameters and about 10 billion activated per inference, with an average time to first token within 3 seconds.
- Training built nearly 100,000 high-quality SWE tasks autonomously, raising training trajectories by about 60% and cutting training steps by about 55%.
NewTimeSpace (newtimespace.com) News: On 15 September 2026, Unisound AI Technology Co., Ltd. (stock code: 09678) published a voluntary announcement launching U2-Flash, a new-generation high-density intelligent model built on enhanced post-training of the U2 general foundation model.
U2-Flash doubled its predecessor's score on the DeepSWE v1.1 coding benchmark, scoring 64.6 and surpassing GLM5.3-Flash and DeepSeek-V4-Pro-0813. Its TerminalBench3.0 score rose to 24.3, above trillion-parameter models such as K3, and its SWE-Bench Pro score reached 61.6, up 10.5 points from the previous generation. Inference cost and end-to-end execution were also optimised, with Agent task iterations cut by 20% to 30%, task execution cycles shortened by 35% and token consumption reduced by 20% to 30%.
U2-Flash uses a sparse mixture-of-experts architecture with about 266 billion total parameters and about 10 billion activated per inference, an average time to first token within 3 seconds and peak output throughput of up to 300 tokens per second. It unifies coding, agentic, mathematical reasoning and instruction-following capabilities in one set of weights, features implicit thinking and continuous-state reasoning, offers four controllable levels of reasoning intensity, and has completed systematic adaptation to mainstream domestic computing platforms.
On training, the model takes part in its own training through an autonomous closed loop, building nearly 100,000 high-quality SWE tasks on its own, raising training trajectories by about 60% and cutting training steps by about 55%. All adjustments are made within manually defined sandbox environments and validation standards, with complete and rollback-enabled records retained.
NewTimeSpace Disclaimer: All content herein is the original work of NewTimeSpace. Any reproduction, reprinting, or use of this content in any other manner must clearly indicate the source as "NewTimeSpace". NewTimeSpace and its authorized third-party information providers strive to ensure the accuracy and reliability of the data, but do not guarantee the absolute correctness thereof. This content is for reference only and does not constitute any investment advice. All transaction risks shall be borne by the user.
- PINE TECH (01079.HK): rights issue oversubscribed by about 13.5%, raising HK$105 million net for memory and server inventory
- HIGHTIDE-B (02511.HK): Terminates Acting-in-Concert Agreement
- Going Global Weekly: Innovative Drug Data and Regulatory Qualifications Go Global in Clusters as Tech IPOs See Pricing and Terminations Side by Side-20261030
- HK Stock Connect Daily Report: EVERG SERVICES Falls 17.72% as PAG Agrees to Acquire a Controlling Stake for HK$4.026 Billion - 20261030
- YUK WING GP (01536.HK): Rights Issue Done; 1.368 Billion Shares, HKD 134.9m
