Two Months After Founding, Shanghai AI Lab's Former Model Chief Raises a Nine-Figure Yuan Angel Round for Self-Improving Base Models

A two-month-old Shanghai startup has landed one of China's larger angel rounds of the year. Superevolution Intelligence (超进化智能), founded in August in Zhangjiang, Pudong, says it has raised several hundred million yuan — a nine-figure sum in yuan terms, with the exact figure undisclosed — in a round led by BlueRun Ventures China, with SenseTime-affiliated Guoxiang Capital, Legend Star, the Inno Angel Fund and the Shuimu Tsinghua Alumni Seed Fund participating. A Chinese startup-data weekly ranked the round among the top 5% of angel rounds by size this year.

The bet is recursive self-improvement — RSI — at the model layer. Where most of today's agent companies iterate prompts, tools and harness code around a frozen model, Superevolution says it wants base models that internalize new capabilities through parameter updates, so improved skill travels with the weights instead of living in an external framework. The company describes a five-level ladder, from updating research artifacts and runtime experience assets, through self-iterating harness and agent code, to model-layer self-evolution, and finally "meta-evolution," where the improvement mechanism itself is iterated by AI.

None of this would carry much weight without the founder. Chen Kai studied at Tsinghua and took his PhD at the Chinese University of Hong Kong's Multimedia Lab under the late Tang Xiaoou and Lin Dahua — the lab that produced several SenseTime co-founders. He was a core lead of OpenMMLab, the open-source computer-vision stack with more than 140,000 GitHub stars, joined SenseTime as an R&D director in 2019, and from 2022 ran the large-model center at the Shanghai AI Laboratory, where he led the Intern model family, oversaw end-to-end training of a trillion-parameter science model on a 10,000-GPU cluster, and built the OpenCompass evaluation suite now widely used across Chinese labs. His team says it released Intern-S1-Pro — billed as the world's first trillion-parameter science foundation model — this past March.

The roadmap, per Chinese tech-media interviews: start from post-training with models of several hundred billion parameters, graduate to mid-training, and attempt full pretraining once resources allow, inside a closed loop where the model, the agent harness and the evaluator co-evolve.

The investor list reads like a coalition of everyone with a stake in that track record. BlueRun is betting on model-layer depth over the harness-layer iteration it sees across the agent market. SenseTime's Guoxiang Capital is, in effect, an MMLab lineage fund backing a second-generation founder. The Tsinghua alumni fund is the personal-confidence vote. All are entering at the earliest stage, which tells you the diligence was about the team, not the product.

There is, as yet, no model, benchmark or demo to evaluate — the framing above is the company's own account via Chinese tech media. But the timing lands in a moment when recursive self-improvement has moved from fringe to roadmap: Anthropic co-founder Jack Clark recently said it could begin in the near term, and SemiAnalysis' new GDPx benchmark includes an entire RSI domain where GPT-6 Astra leads the field.

The broader signal is about China's post-DeepSeek startup wave: rather than competing head-on with hyperscalers for general chatbot supremacy, a growing set of well-credentialed teams is picking deep, narrow problems — science foundation models, world models, and now self-improving ones. A lab run by the person who built the country's de facto evaluation standard choosing evaluation as part of its own flywheel is a credible setup. Whether the closed loop actually runs will show up, as always, in the models.