What happened
Chaoyan Intelligence, a startup founded by 27-year-old Tsinghua University assistant professor Chen Yongchao, has released Apex-35B-V1.0, a 35B parameter AI model. The model is built upon the Qwen3.6-35B-A3B base and incorporates proprietary datasets and research trajectories. The company claims the model utilizes Recursive Self-Improvement (RSI) to allow the AI to participate in its own research and development cycles. According to 36Kr, the model is designed to handle complex scientific research, agentic tool-calling, programming, and reasoning tasks.
Chaoyan Intelligence, founded in June 2026, announced the release of Apex-35B-V1.0 on October 22. The model is a 35B parameter system that utilizes a MoE-like approach to balance general-purpose tasks with specialized research capabilities. The company reports that the model has been trained using a combination of the Qwen3.6-35B-A3B base and proprietary data focused on research trajectories.
The core innovation highlighted by the company is the integration of Recursive Self-Improvement (RSI). This methodology allows the model to participate in experimental design, result analysis, and the application of verified findings back into the R&D pipeline. The company claims this approach has allowed them to achieve competitive results in 21 tests, including FrontierScience and PaperBench, where it reportedly outperformed significantly larger models.
The company has secured approximately 400 million yuan in funding over its first four months of operation. The model is currently accessible via Hugging Face, though specific commercial licensing terms or enterprise deployment costs were not detailed in the report.
Why it matters
The release of Apex-35B-V1.0 highlights a growing industry trend toward 'Recursive Self-Improvement' (RSI), where AI models are tasked with contributing to their own R&D processes to accelerate innovation. By achieving performance metrics that reportedly rival or exceed much larger models (500B-800B parameters) in specific scientific and reasoning benchmarks, the model demonstrates that specialized training and architectural choices can potentially bridge the gap between mid-sized models and massive, compute-heavy systems. This development is significant for researchers and developers seeking high-performance, specialized AI capabilities that can operate with greater efficiency than traditional large-scale models.
The shift toward RSI represents a strategic attempt to overcome the diminishing returns of simply scaling parameter counts. By enabling models to assist in their own development, companies like Chaoyan Intelligence aim to create a self-accelerating cycle of technological progress.
The reported performance of Apex-35B-V1.0 in benchmarks like the Berkeley Leaderboard (BFCL V4) and Humanity's Last Exam (HLE) suggests that mid-sized models, when optimized for specific reasoning and research tasks, can achieve parity with or surpass models an order of magnitude larger in parameter count. This has practical implications for organizations looking to deploy high-capability AI in resource-constrained environments.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Observers should monitor the practical, real-world application of Apex-35B-V1.0 beyond scores, particularly its performance in actual scientific research and software engineering environments. As the model is available via Hugging Face, independent verification of its capabilities in reproducing research papers and executing complex agentic tasks will be critical to validating the company's claims. Additionally, the long-term efficacy of the RSI methodology in sustaining model improvement without degradation remains a key area for technical scrutiny.
Independent validation of the model's performance on the cited benchmarks is necessary, as internal benchmarks can sometimes be optimized for specific test sets. The community will likely focus on whether the model maintains its performance when applied to novel, unseen research problems.
The sustainability of the RSI approach is a major unknown. While the company claims success in early iterations, the industry will watch to see if this method leads to 'model collapse' or other degradation issues over multiple generations of self-improvement.
The company's ability to scale its operations and maintain its development pace will be tested as it moves from its initial release to broader enterprise adoption.