HyperVolve Intelligent has closed an angel round in the several hundreds of millions of yuan. BlueRun Ventures China led, with SenseTime Guoxiang Capital, Legend Star, Inno Angel and a Tsinghua alumni seed fund participating. Lansong Capital is long-term strategic financial adviser. The company began operating in late August.
Two months and no product is, on its own, unremarkable in Chinese AI financing. What makes this round worth examining is the specific technical bet and the specific person making it.
The CV behind the cheque
Chen Kai took his undergraduate degree at Tsinghua and his doctorate at the Chinese University of Hong Kong’s Multimedia Laboratory. He was a research director at SenseTime and later headed the large model centre at Shanghai AI Laboratory.
Three items on that record explain the round. He created OpenMMLab, the computer vision open-source framework that has passed 140,000 GitHub stars and became default infrastructure for a generation of vision researchers. He led work on the Intern model series, including full-process training of a trillion-parameter scientific model on a ten-thousand-GPU cluster. And he built OpenCompass, the evaluation suite that a large share of Chinese model developers now use as a reference.
In a field where many founders have published well and shipped little, the distinguishing fact is that Chen has run the entire training pipeline at that scale. The number of people in China who have done so is small enough to be counted, and the market prices that scarcity directly.
What recursive self-improvement means here
HyperVolve’s stated goal is a research brain capable of meta-research: given an unfamiliar technical field, it formulates its own questions, generates hypotheses, designs experiments, validates results and iterates.
The company lays out four conditions it believes have now been met. Model reasoning in specialised domains has reached the level where a model can propose the next step rather than only assess a result. Coding and agent capability allow tasks that run for hours. Agentic reinforcement learning gives a tractable path for models to update from interaction with an environment. And automated evaluation has become robust enough that the reward signal can be trusted.
Chen’s summary of the shift is compact. Coding, he argues, was the first professional capability where large models closed a large-scale iteration loop. Research will be the next.
The architecture follows from that. HyperVolve separates self-improvement into five layers: the artefact layer, where research output corrects itself automatically; the experience layer, where memory and runtime skills are updated; the harness layer, where the agent’s own code iterates; the model layer, where new capability is internalised through parameter updates; and the evolution-mechanism layer, where the method of improvement is itself handed to the model.
The company’s own assessment is that most of the field is working in the top three layers, because the harness is tractable and produces visible results quickly. Only parameter-level updating counts as real self-evolution, and that is also the layer where stability, evaluation integrity and transfer across domains remain unsolved.
The crowded field next door
HyperVolve is not alone in this direction, which is part of why the round closed quickly. Chaoyan Intelligence launched in July and raised close to RMB 400 million by September, also targeting scientific research. Chuguang Yueqian and Xinyuanti are pursuing what they call physical recursive self-improvement, applying the same logic to robots that improve through real and simulated tasks.
The shared premise is that the next capability gain comes not from a better base model but from a system that improves the base model. The differences are in where the loop closes: in software, in a laboratory, or in a physical machine.
At the point of announcing the round, HyperVolve had released no product and disclosed no revenue. The model layer, the one that matters most, has not been demonstrated. Whether parameter self-updates remain stable, whether an evaluator can resist a model that learns to satisfy it rather than to solve the task, and whether research ability transfers across fields are all open.
Capital has decided to fund the question rather than the answer. That is a defensible position at an angel stage, and it is also the position every deep technology bet occupies before its first result. The round buys compute and researchers. It does not buy evidence, and the next twelve months are where the evidence has to appear.
Editor’s note: This is an adapted translation of the original Gasgoo report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.gasgoo.com/apps/e060bad9eca91db7f01403fc7120041f/robot/news/70476230-seeds.