Today, Unisound officially launches U2-Flash.

Rather than being a lightweight efficiency version of U2, U2-Flash represents a new-generation intelligent model designed for real-world productivity.

As a high-density intelligence model, U2-Flash leverages systematic post-training to unlock advanced reasoning and autonomous execution capabilities. It combines the robust capabilities of a flagship model with Flash-level responsiveness, delivering a stronger, faster, and more cost-efficient solution for real-world productivity scenarioses.

Compared with many other large model companies, Unisound’s unique advantage lies in more than ten years of experience accumulated across real-world application scenarioses. Long-term business practice has given Unisound deeper scenario understanding and richer data accumulation, enabling U2-Flash to be designed from the beginning around the needs of real productivity.

U2-Flash follows two core principles in capability development:
First, the model deeply participates in its own training loop evolution — from generating training data and analyzing trajectories to inspecting and repairing training systems. This represents an early practice toward Recursive Self-Improvement (RSI) and serves as a central thread throughout the technical design.
Second, the model performs deep thinking in latent space (Latent Reasoning), while exposing controllable thinking intensity interfaces to users instead of treating reasoning as an unobservable black box.

Architecturally, U2-Flash adopts a Sparse Mixture-of-Experts (MoE) design, with approximately 266B total parameters but only around 10B parameters activated during each inference process. It integrates capabilities including coding, Agent tasks, mathematical reasoning, and instruction following into a unified set of weights, achieving flagship-level task completion quality with far fewer activated parameters.

01 Stronger: Capability Breakthroughs Comparable to Flagship Models

In real business scenarioses, model competition has shifted from simply answering questions to completing complex tasks.

Compared with the previous-generation U2 model, U2-Flash delivers comprehensive improvements across coding, Agents, advanced reasoning, and office automation scenarioses, achieving strong performance across multiple authoritative benchmarks.

In DeepSWE v1.1, a benchmark representing coding capability, U2-Flash achieves 64.6 points, surpassing models including GLM5.3-Flash and DeepSeek-V4-Pro-0813. In TerminalBench 3.0, it reaches 24.3 points, exceeding trillion-parameter-scale models such as K3. In SWE-Bench Pro, it achieves 61.6 points, improving significantly over the previous generation.

These results demonstrate that with less than 4% of total parameters activated, U2-Flash delivers comprehensive capabilities comparable to mainstream flagship models.

02 Faster: From Fast Generation to Fast Task Completion

For complex workflows, token generation speed only reflects throughput. End-to-end task completion time is the true measure of efficiency.

U2-Flash accelerates the entire process from initial response to final task delivery. Its average time to first token (TTFT) is controlled within 3 seconds, with peak output speed reaching 300 tokens/s. Compared with U2, U2-Flash reduces Agent task iteration steps by 20%-30% and shortens task completion cycles by 35%, enabling faster response, generation, and completion.

03 More Efficient: Making Every Token More Valuable

For frequent real-world tasks, U2-Flash further improves token utilization efficiency by reducing ineffective exploration and redundant outputs, allowing more tokens to directly contribute to solving problems.

Compared with U2, U2-Flash can reduce token consumption by 20%-30% during complex task execution, lowering the cost of each task while improving efficiency.

This efficiency is also reflected at the architecture level. Approximately 10B activated parameters mean each inference only needs to activate a small portion of the total model parameters, allowing inference costs to scale with activated parameters rather than total parameters.

Meanwhile, U2-Flash integrates coding, Agent tasks, reasoning, tool calling, and office processing into one unified model, reducing the complexity and cost associated with maintaining multiple specialized models.

04 High Performance on Domestic AI Computing Platforms

Advanced intelligence should not be limited to a single computing platform. It should be able to run efficiently and stably across diverse heterogeneous computing environments.

U2-Flash continuously optimizes adaptation for major domestic AI computing platforms, covering model architecture, core operators, inference frameworks, and cluster scheduling. Through hardware-software co-optimization, it reduces deployment costs while improving inference efficiency and resource utilization.

In practical services, U2-Flash continues improving throughput, latency, concurrency, and cluster scalability on domestic platforms, narrowing the gap with mainstream GPU solutions. In some scenarioses, performance approaches NVIDIA GPU solutions.

05 Post-training × RSI Driven: Unlocking High-Density Intelligence Through Autonomous Evolution

Post-training is the core engine behind U2-Flash’s high-density intelligence.

For complex tasks, simply increasing model scale is no longer sufficient. U2-Flash uses systematic post-training as a key approach to strengthen understanding, planning, and execution capabilities.

U2-Flash establishes an autonomous closed-loop evolution mechanism. The model participates deeply in task generation, trajectory analysis, and error correction. Through comparisons between stronger and weaker models, it identifies critical actions and continuously improves weak areas without requiring massive human annotation.

Unisound’s more than ten years of industry experience has accumulated extensive real-world scenario data and high-quality annotation capabilities, providing an important data foundation for post-training.

This closed loop represents U2-Flash’s first step toward Recursive Self-Improvement (RSI). The model can participate in generating training data, analyzing execution trajectories, and inspecting training systems, becoming not only an object being trained but also a participant in the training loop.

However, this self-improvement remains bounded, verifiable, and traceable. Every autonomous adjustment occurs within human-defined sandbox environments and validation standards.

U2‑Flash has officially launched on the Unisound MaaS Platform. Experience it now: http://maas.yhjs818.com/models/u2-flash