Gasgoo Munich- HyperVolve has closed an angel round totaling hundreds of millions of yuan, Gasgoo Embodied Intelligence has learned.
The fresh capital will mainly fund computing and training infrastructure, along with expanding the core R&D team.
Founded and launched in late August, HyperVolve focuses on developing self-evolving research foundation models. It uses Recursive Self-Improvement (RSI) as its core technical path, pushing models to autonomously sharpen their professional capabilities in unfamiliar scenarios.
How does a startup barely two months old secure hundreds of millions in angel funding?
Image credit: HyperVolve
Betting on "Research is the next Coding"
HyperVolve founder Chen Kai holds a bachelor's degree from Tsinghua University and a PhD from the Multimedia Lab (MMLab) at the Chinese University of Hong Kong. He previously served as research director at SenseTime and head of the Large Model Center at the Shanghai Artificial Intelligence Laboratory.
His track record features three widely cited projects. First, the OpenMMLab computer vision open-source algorithm system he created, which has garnered over 140,000 GitHub stars. Second, the Intern large model he led at the Shanghai AI Lab, where he directed the full-pipeline training of a trillion-parameter scientific model across a 10,000-GPU cluster. Third, the OpenCompass large model evaluation framework he spearheaded, now widely adopted by leading AI companies and research institutions.
Chen has published over 100 papers, with more than 45,000 citations on Google Scholar.
In short, this is someone who has truly mastered 10,000-GPU clusters and trained trillion-parameter models.
HyperVolve’s core team also draws heavily from the Shanghai AI Lab and top tech firms, bringing deep expertise in large-scale foundation model training, evaluation systems, and scientific models.
HyperVolve has set its sights on building a "super researcher brain" endowed with meta-research capabilities. When confronting unfamiliar professional domains, the model should autonomously pose questions, formulate hypotheses, design experiments, self-verify, and iterate. These capabilities are designed to transfer across scientific research, industrial R&D, and AI research.
HyperVolve argues that several prerequisites for the "industrialization of research capabilities" are now falling into place. First, models' professional reasoning now rivals average experts in many fields, evolving from merely analyzing results to proposing next steps. Second, advances in coding and agentic capabilities allow models to execute tasks spanning hours or more. Third, Agentic RL offers a relatively tractable path for models to update parameters through environmental interaction. Fourth, automated and robust evaluation, coupled with rapid progress in domain verifiers, makes feedback and reward signals far more reliable.
Layered together, these four shifts turn self-evolution from a conceptual puzzle into an engineering challenge.
"Coding was the first professional capability where large models closed the loop on large-scale iteration," Chen said. "Research will be next."
RSI has five layers — HyperVolve is going straight for the hardest: the "model layer"
HyperVolve breaks RSI into five layers, from shallowest to deepest. The first is the artifact layer, where auto-research self-corrects its outputs. Next is the experience layer, which continuously updates runtime assets like memory and skills. Then comes the harness layer, iterating the agent system's own code. The fourth is the model layer, where continuous parameter updates truly internalize new capabilities. At the deepest level lies the evolution mechanism layer, handing the methodology of "how to evolve" itself over to AI.
In HyperVolve’s view, most current efforts cluster around the first three layers because the harness layer has a lower barrier and delivers quick wins. Yet the true vessel of intelligence is model parameters; only continuous parameter-level updates constitute genuine self-evolution.
HyperVolve has chosen the fourth layer.
That means the startup isn't betting on a proven, mature direction. It is targeting the hardest entry point — one where industry consensus has yet to form.
To that end, HyperVolve has built a closed loop where the model, the agent harness, and the evaluator co-evolve, while simultaneously constructing an evaluation framework that covers both research capabilities and self-improvement abilities.
Chen believes the two scarcest resources for RSI both depend on foundation model training experience. One is full-pipeline experience training trillion-parameter models across 10,000-GPU clusters — very few have actually done it. The other is the long-term accumulation of research trajectory data and professional task environments, a know-how forged through trial and error.
RSI becomes the new narrative; the track starts to stratify
Zooming out, more companies are placing bets on model self-improvement.
HyperVolve is playing for research, with science as its main battleground. Its meta-research capabilities, however, can transfer to scientific research, industrial R&D, and AI research.
Chaoyan Intelligence, which announced its founding in July and closed nearly 400 million yuan in funding in September, also focuses on scientific research and discovery. Its goal is to evolve AI from an assistive tool into a co-researcher — shifting from accelerating the known to discovering the unknown.
Image credit: NEXFORM
Chuguang Yueqian and NEXFORM are working on Physical RSI, grounding self-improvement in the physical world and robotics, allowing models to iterate through real or simulated tasks.
The shared vision is "making models stronger on their own," but some target research while others target physical embodiment — their entry points and priorities diverge.
A dose of reality: this remains an exceedingly early track.
For HyperVolve specifically, the startup had yet to release a product or disclose revenue as of the funding announcement. Moreover, the "model layer" of RSI is the most ideal but also the hardest. The stability of parameter self-updates, whether models can "game" the evaluator, and whether research capabilities can transfer across domains all remain unproven.
By that measure, capital is willing to place dense bets on the self-evolution thesis — but companies that truly close the "research loop" will need time and compute to prove it.
About 'Seeds':
Gasgoo Auto's 'Seeds' column aims to build a service platform linking startups, supply-chain partners, investors, and local governments to deeply empower the industry ecosystem. Since its launch, the column has been dedicated to unearthing good companies, technologies, and business models that inspire and lead amid the intelligent transformation wave — fueling the growth of industrial innovation.
