Tencent's New AI Trick Lets Models Teach Themselves in a Loop
Tencent has unveiled a training technique it calls Terminal-RSI, which allows a large language model to improve itself by generating its own practice problems and then solving them. The company described the method as a way to give models a kind of self-directed learning loop, reducing reliance on human-written training data. The announcement, reported by tech outlet TMTPost, points to a growing focus among Chinese AI developers on making models more capable without simply feeding them more text.
The core idea is that a model can act as both student and teacher. It proposes tasks, attempts them, checks the results against verifiable outcomes, and learns from its mistakes. Over repeated cycles, the model’s performance improves on the tasks it invents. Tencent calls this a small-scale version of recursive self-improvement, a concept that has long fascinated AI researchers but remains difficult to implement reliably. The company’s framing suggests it sees the approach as a practical path to stronger reasoning skills.
Tencent’s move comes as China’s AI industry splits into distinct camps. DeepSeek, the Hangzhou-based lab that startled the world with its low-cost models, has focused on efficiency and open releases. Kimi, developed by Moonshot AI, has staked out long-context document processing. Tencent, a Shenzhen-based internet giant known for WeChat, has been building its Hunyuan family of models and integrating AI across its consumer and enterprise products. The Terminal-RSI work appears aimed at closing the gap in reasoning, where American models have often led.
The timing is notable. DeepSeek recently began charging for web search in its API, a sign that Chinese AI firms are moving toward sustainable business models after a period of aggressive price competition. Tencent’s research, by contrast, is about capability rather than cost. If models can improve themselves with less human supervision, the economics of AI development could shift, making it cheaper to reach higher performance and potentially accelerating the pace of releases.
Still, self-improving models raise questions. A model that generates its own training data can reinforce its own flaws, a problem known as model collapse. Tencent’s method relies on tasks with verifiable answers, which limits the domains where it can work. Creative writing, open-ended reasoning, and real-world planning are harder to check automatically. The company has not said which products or future Hunyuan models will use the technique.
The broader context is a Chinese AI market that has become both crowded and creative. Startups and giants alike are experimenting with agentic systems, search tools, and personal assistants. Tencent’s Terminal-RSI fits that pattern: a research advance that could feed into products, from cloud services to consumer apps. For global observers, it is a reminder that China’s AI progress is not only about catching up on raw scale but also about finding new ways to train and deploy models.
What happens next depends on results. Tencent has published little detail on benchmarks, and independent verification will matter. But the direction is clear. Chinese labs are no longer just following; they are proposing their own answers to the hardest problem in AI, which is how to make models smarter without running out of data or money.