Apex Intelligence has raised nearly US$50 million across angel and angel-plus funding rounds as it develops foundation models designed to improve themselves through recursive self-improvement (RSI). Founded in June 2026, the company is targeting scientific research rather than conventional AI assistance, with a system intended to generate hypotheses, conduct experiments, validate results and iteratively improve its own capabilities.
The next phase of AI development may depend less on making models better at answering questions and more on making them capable of generating and testing new ideas.
That is the premise behind Apex Intelligence, a newly founded AI company developing what it describes as self-evolving foundation models. The company has completed angel and angel-plus funding rounds totaling nearly $50 million, providing capital for foundation-model development, computing infrastructure and research data systems.
The angel round was co-led by IDG Capital, LinkX Capital and XtalPi, with participation from Decent Capital, SEE Fund, Monad Ventures and Winsoul Capital. The angel-plus round was co-led by Zhongguancun Science City Fund, SCGC and Shanghai Engine Fund.
Apex’s technical focus is recursive self-improvement, or RSI. Instead of relying entirely on humans to determine what a model should learn next, the company’s approach is designed to have models participate in an iterative research loop involving hypothesis generation, experimentation, validation and further improvement.
That ambition places Apex in a different part of the AI infrastructure landscape from products primarily designed to make existing knowledge easier to access or automate established workflows.
The company wants its models to function as AI researchers capable of contributing to scientific and technical discovery. Its proposed applications include chip design and simulation, molecular research and quantitative strategy development—areas where experimentation can be expensive and where outcomes can be measured against technical objectives.
Apex says its early work has focused on two areas: AI for AI and AI for Math.
For AI research, the company says its system has autonomously generated research it considers suitable for submission to leading AI conferences and has produced improvements across areas including pre-training decisions, training strategies and GPU-kernel optimization. Apex also says its systems have set or approached leading results on benchmarks and research projects including SimpleTES, NanoChat Autoresearch, GPUMode TriMul and MLS-Bench.
Those claims are significant but should be viewed as company-reported results rather than independent evidence that the system has reached human-researcher equivalence.
Apex also reports results in mathematics, including a complete proof of what it calls the majorization conjecture, alongside work in optimization theory and geometric topology. As with its AI research results, the broader significance of these claims will depend on independent verification, publication and reproducibility.
The company’s strategy is built around model training. Apex says mid-training and post-training will be used to teach models to generate research ideas and autonomously execute iterative research processes.
That puts training methodology and research trajectory data alongside model architecture and compute as important parts of its proposed AI stack. The company says it plans to build infrastructure for high-quality research trajectory data while continuing to invest in foundation models and compute.
The approach reflects a broader shift in AI research toward systems that can use tools, write and execute code, evaluate outputs and refine strategies rather than simply generate a single response. Agentic AI systems already automate portions of software development, data analysis and scientific workflows, but Apex is targeting a more ambitious objective: making the research process itself part of the model’s operating loop.
Apex was founded by Yongchao Chen, an assistant professor at Tsinghua University’s School of Artificial Intelligence. The company says its team includes researchers with backgrounds at Tsinghua University, Peking University, Harvard and MIT, while Chen has conducted research at Google DeepMind, Microsoft and the MIT-IBM Watson AI Lab.
The company is also seeking additional researchers. From September 17 to September 23, Apex plans to visit Harvard, MIT, Yale and Boston University to engage with students and academic research groups around recruitment and research collaboration.
Its longer-term vision is considerably more ambitious than today’s AI assistants. Apex says it wants to develop a scalable network of AI researchers whose collective capabilities could eventually match or exceed those of large numbers of top human scientists.
That remains a long-term company objective rather than an established capability. The more immediate technical question is whether recursive improvement can produce measurable, reproducible gains across increasingly difficult research problems.
If it can, the implications extend beyond another foundation-model release. AI systems capable of reliably improving research strategies could become part of the infrastructure used to develop new algorithms, materials, drugs, chips and other technologies.
Market Landscape
Foundation-model development is increasingly expanding from static training toward agentic systems, automated experimentation and AI-assisted research.
Companies and research organizations are exploring systems that can write code, operate tools, run experiments and evaluate their own outputs. Recursive self-improvement represents a more ambitious extension of this trend because the system is expected to improve aspects of its own problem-solving process rather than simply execute a predefined workflow.
Apex Intelligence is entering a competitive environment that includes major AI research organizations such as Google DeepMind, OpenAI, Microsoft Research and NVIDIA, as well as university research groups and emerging AI startups.
The central technical challenge is verification. A self-improving system needs reliable evaluation mechanisms so that improvements are genuine rather than artifacts of benchmark optimization. For scientific applications, reproducibility, independent validation and the ability to distinguish novel discoveries from errors will become particularly important.
Top Insights
- Apex Intelligence has raised nearly $50 million to develop foundation models based on recursive self-improvement and autonomous research workflows.
- Its proposed AI researcher architecture combines hypothesis generation, experimentation, validation and iteration rather than limiting models to human-directed assistance.
- The company reports early results across AI research and mathematics, although its strongest capability claims remain subject to independent verification.
- Apex plans to invest in foundation models, compute and research-trajectory data as core infrastructure for self-evolving AI systems.
- The company’s long-term strategy targets scientific discovery across fields including chip design, molecular research and quantitative strategy development.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
