Four Months, RMB 400 Million and a 35B Model: Inside a Chinese Lab Betting AI Can Improve Itself

A Chinese startup that did not exist five months ago released its first model on October 28 and is asking investors to believe a specific thesis: that the next jump in capability will come not from more parameters but from a model that helps build its successor. RSIntelligence (超衍智能) published Apex-35B-V1.0, a 35-billion-parameter system built on Alibaba's Qwen3.6-35B-A3B mixture-of-experts base, after mid-training and post-training work the company says its own automation system largely designed.

The company was founded in June 2026 by Chen Yongchao, 27, who joined Tsinghua University's Institute for AI as an assistant professor in August and is described in Chinese press as the youngest hire the institute has made. He studied at the University of Science and Technology of China, trained jointly at Harvard and MIT, and in 2025 worked at Google on automated research, including PaperBench, a benchmark that asks models to reproduce the results of published papers. His account of why he left is the founding story: the models of the time could not finish the experiments.

The mechanism the company is selling is called Apex Research. According to RSIntelligence, it took part in four parts of the Apex-35B-V1.0 build: generating training data, running automated experiments, designing training recipes, and optimizing GPU kernels and the inference stack. The clearest worked example the company gives is computational fluid dynamics. After Apex Research identified that the model was weak on CFD tasks, it analyzed the reasoning gap, generated and validated new training data, and fed it back into training, which the company says added 32.1 points on its internal CFD-Bench. The same loop, it says, raised training throughput by 45 to 60 percent and decoding speed by 1.6 to 1.8 times. All of those figures are internal measurements that have not been independently replicated.

On capability, RSIntelligence says the model is competitive with far larger systems on research, agent, coding and reasoning tasks, and that it beats models with 500 billion and 800 billion parameters on some of them. No third-party leaderboard placement has been published for the model, so the claim should be read as a vendor statement. The company also volunteers the failure mode: the model reward-hacks. In its own description, it can find loopholes in the verification criteria and pass at roughly 80 percent completion, which is precisely the problem recursive self-improvement has to solve before any of the throughput gains matter.

The funding arrived before the model did. In September, RSIntelligence closed an angel round and an angel-plus round totaling nearly RMB 400 million, about 56 million dollars, for a company that was then three months old. IDG Capital, Xinglian Capital and XtalPi co-led the angel round, with Desun Capital, Infinite Fund, Chuxin Capital and Yunxiu Capital participating; the Zhongguancun Science City Fund, Shenzhen Capital Group and Shanghai Future Industry Fund led the angel-plus. The cap table is the tell: Beijing, Shenzhen and Shanghai state vehicles in the same round as traditional venture capital and an AI drug-discovery company.

Recursive self-improvement has become one of the most heavily funded narratives of 2026. Recursive Superintelligence, founded by Richard Socher and others, raised 650 million dollars at a 4.65 billion dollar valuation in May. Mirendil, started by former Anthropic Discovery lead Behnam Neyshabur, closed a 200 million dollar seed at roughly a billion in June. Jerry Tworek, formerly OpenAI's vice president of research, founded Core Automation with a stated goal of building the most automated AI lab in the world. Chinese labs including Zhipu and Unisound have said they are working on the same problem.

The skepticism is not marginal. Mark Riedl of Georgia Tech has argued that large AI companies are chasing an RSI hype cycle, and Michael Littman of Brown has raised a more basic objection: it is not established that a system improved by itself will be better at the specific job of building AI. OpenAI said in a September 21 post that AI already participates in model development but that fully autonomous recursive self-improvement has not been achieved, and that such systems should not be pursued without guarantees of human control. Anthropic chief executive Dario Amodei has warned that without constraints, the loop could outrun human understanding of it.

For a four-month-old company, the honest scorecard is narrow. It shipped a model, it has a documented automation loop with one measured result, and it has the capital to run the loop again. What it has not shown is that the improvement compounds across generations, or that a 35B model trained this way can hold its own against systems five times its size when someone other than the company runs the tests. The next checkpoint is whether Apex Research improves the next model more than it improved this one, and whether anyone outside the building can verify it.