On 15 September 2026, Unisound AI Technology Co., Ltd. announced the launch of U2-Flash, a new-generation high-density intelligent model positioned as an upgrade to the U2 foundation model.
U2-Flash posted a DeepSWE v1.1 score of 64.6—twice the previous model’s performance and ahead of peers such as GLM5.3-Flash and DeepSeek-V4-Pro-0813. TerminalBench 3.0 reached 24.3, surpassing trillion-parameter models including K3, while SWE-Bench Pro climbed 10.5 points to 61.6.
Operational metrics also improved materially. Iteration steps for Agent tasks fell by 20%–30%, the task execution cycle shortened by 35%, and token consumption declined 20%–30%, collectively lowering end-to-end execution costs.
The model adopts a sparse Mixture-of-Experts architecture with 266 billion parameters, activating roughly 10 billion parameters per inference. Average time to first token remains below three seconds, and peak throughput reaches 300 tokens per second. Multi-domain capabilities—programming, Agent operations, mathematical reasoning, and instruction following—are unified within one weight set, complemented by a four-level controllable reasoning interface that provides verifiable reasoning chains at higher intensities.
Training incorporates an autonomous closed-loop mechanism, marking Unisound AI’s first step toward Recursive Self-Improvement. Innovations such as asynchronous Agent RL, online policy distillation from multiple self-trained teacher models, and adaptive task generation lifted effective training trajectories by about 60% and reduced training steps by roughly 55%. All self-adjustments occur within a sandbox governed by predefined validation standards and full rollback records.
Management highlighted that U2-Flash moves the company from “human-led model training” to “model-assisted evolution,” enabling continuous self-validation and self-enhancement while maintaining safety and traceability. The model has already been adapted to mainstream domestic computing platforms, positioning Unisound AI for broader commercial deployment of high-density intelligence solutions.