Yunzhisheng Releases U2-Flash, a Domestic RSI Model That Lets Flash Beat Flagship

U2-Flash uses a sparse mixture-of-experts architecture with about 266 billion total parameters and activates about 10 billion parameters per inference, less than 4 percent of the total. Compared with U2, it generates 2.1 times faster, shortens agent task completion time by 35 percent, and reduces task iteration steps and token consumption by 20 percent to 30 percent. On benchmarks cited in the report, DeepSWE v1.1 rose from 32 to 64.6, TerminalBench 3.0 jumped from 2.7 to 24.3, and SWE-Bench Pro reached 61.6, up 10.5 points from the previous generation. The report said some results enter the range of models with trillion-level parameters.

Yunzhisheng has placed U2-Flash on its MaaS platform for online trial and API access. The API is compatible with both OpenAI and Anthropic protocols, and the company provides access instructions for tools including Claude Code, Trae, Cursor, and Cline. U2-Flash offers four thinking levels: none, low, high, and max. The model is currently offered at a limited 60 percent discount, with input priced at 0.6 yuan per million tokens, output at 1.2 yuan per million tokens, and cache hits at 0.12 yuan per million tokens. From Sept. 15 to Sept. 30, users can claim 100 million free tokens.

QbitAI tested U2-Flash in three scenarios. In the first, it planted three bugs in a fictional ExpenseFlow expense-reporting project: a rejected reimbursement was included in the reimbursed total, a Beijing-time record submitted after midnight on Aug. 1 was stored in UTC and pushed back into July, and dates shifted months during CSV export. Without clues, U2-Flash found all three issues in 7 minutes 6 seconds and delivered an acceptance report. In the second, a ProfileSync 2.0 project contained conflicting documentation, including a README command to run python app.py even though the file did not exist. U2-Flash identified the real fault in 3 minutes 32 seconds, fixed it, and expanded tests from 1 to 7, all of which passed. In the third, it handled 14 files in four formats for a simulated product launch and produced a DOCX file containing a 1,200- to 1,500-word release, a source table for key facts, a list of material conflicts, and five pre-release questions. It completed the task in 8 minutes 36 seconds, used the latest Oct. 18 date instead of an older Oct. 8 date, and updated an older 81.2 percent figure to 78.4 percent, noting it was an internal test.

The company attributes the model's performance to data and a post-training closed loop. Yunzhisheng says more than a decade of industry work has given it large volumes of real-world data and high-quality annotation capacity. The loop has three main parts. First, autonomous task generation creates new tasks based on current weaknesses; the system has built nearly 100,000 high-quality SWE tasks across mainstream programming languages. Second, asynchronous agent reinforcement learning distributes tasks to multiple workers in separate sandboxes, marks key actions, and traces final success or failure to decisive steps; the company says this raises effective training trajectories per unit time by about 60 percent. Third, multi-teacher online policy distillation trains specialist teacher models for math, code, and agent tasks, then scores the student model token by token; this reduces the training steps needed for equivalent capability by about 55 percent. The loop also extends to the training system itself, with U2-Flash participating in machine inspection, fault location, repair, and task restart. Yunzhisheng says average recovery time for training faults is about one-third of manual intervention.

The release comes as the broader AI industry debates what constitutes genuine RSI. The report traced the idea to statistician I.J. Good in 1965 and computer scientist and science fiction writer Vernor Vinge in 1993. Anthropic's definition is strict: AI must autonomously design and develop a successor system stronger than itself. OpenAI's observation range is wider, counting a model that accelerates AI research and improves its own capabilities as AI self-improvement, while its Critical level requires fully automated AI R&D or compressing a model generation upgrade cycle to one-fifth. A recent survey paper titled 'The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement' by researchers from Shanghai Jiao Tong University, Tsinghua University, and Shanghai AI Lab laid out a taxonomy starting at L1 based on how much a model takes over the self-improvement loop.