Editorial illustration for China's AI Catch-Up Tests Limits of Western Tech Controls
China's AI Models Match Top US Systems on Benchmarks
China's AI Catch-Up Tests Limits of Western Tech Controls
Kimi K3 and GLM-5.3, the latest releases from Moonshot AI and Zhipu, now sit within striking distance of the top American models on standard benchmarks. That's the headline finding in the fourth issue of THE DECODER's "Frontier Radar," a recurring analysis published for subscribers. Western labs have pointed to distillation, the practice of training on outputs from rival models, as an explanation for how Chinese developers closed the gap so fast.
The editorial team says there's real evidence behind that accusation. But they also argue it barely matters: whether the catch-up came through distillation or independent engineering, the result is the same. A raw model lead, once the thing US labs leaned on to justify their valuations and their infrastructure spending, no longer holds up as a defensible position.
The new issue, the longest in the series so far, traces that shift back to DeepSeek R1's debut a year and a half ago, when a Chinese lab matched OpenAI's o1 and wiped billions off markets in days. From there it asks what Western labs can still claim as an advantage, and why Europe, in the middle of all this, is falling behind on two fronts at once.
Kimi K3 and GLM-5.3 are now within striking distance of the best US models. Western labs blame distillation, and there's real evidence for it. But guilty or not, the conclusion is the same: a model lead can't be defended.
Why this matters
For anyone building on top of frontier models, the distillation debate is a distraction from the actual signal: Kimi K3 and GLM-5.3 are close enough to the best US systems that the gap is now a rounding error, not a moat. Whether Chinese labs got there by skimming outputs or by independent engineering doesn't change what founders and researchers need to plan around. Either way, a model advantage that can be queried through an API has an expiration date.
We'd tell developers to stop pricing in a permanent capability gap when choosing between US and Chinese model providers, and start looking at what actually resists copying: proprietary data pipelines, distribution, compute contracts, applications wired into workflows nobody can scrape. Export controls and chip restrictions still matter, but this issue is a reminder that policy tools built to slow down training runs don't do much once a competitor model is simply good enough to use. The real fight ahead isn't over who has the smartest model this quarter.
It's over what happens after everyone does.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives OpenAI Fixes Codex Bug That Deleted User Files Without Permission
From Archives Coders Find Workarounds to Claude's Invisible Watermarks
Common Questions Answered
How close are Chinese AI models like Kimi K3 and GLM-5.3 to top American models on benchmarks?
According to THE DECODER's Frontier Radar analysis, Kimi K3 and GLM-5.3 from Moonshot AI and Zhipu are now within striking distance of the best US models on standard benchmarks. The performance gap has narrowed to the point where it represents a rounding error rather than a significant competitive moat.
What is distillation and how has it contributed to Chinese AI development?
Distillation is the practice of training AI models on outputs from rival models, allowing developers to accelerate their own model development. Western labs have pointed to distillation as a key explanation for how Chinese developers closed the performance gap with American models so rapidly.
Why does the method Chinese labs used to improve their models matter less than the actual performance gap?
Whether Chinese developers achieved their improvements through distillation or independent engineering, the practical outcome is the same: the performance advantage is now negligible. For developers and researchers, what matters is that a model advantage accessible through an API has a limited lifespan regardless of how it was achieved.
What does THE DECODER's Frontier Radar report reveal about the sustainability of Western AI leadership?
The Frontier Radar analysis demonstrates that a model lead cannot be defended long-term, as evidenced by how quickly Chinese labs have caught up to American counterparts. The report suggests that any technological advantage in frontier models will eventually erode once the capabilities become accessible through APIs.
Further Reading
- GLM-5.3: How Chinese labs keep stride with the frontier - Interconnects
- Reading Zhipu's GLM-5.3 results past the headline number - Artificial Intelligence News
- Kimi K3: second only to Fable 5 on AA-Briefcase - Artificial Analysis
- Kimi K3, and what we can still learn from the pelican ... - Simon Willison’s Weblog
- Chinese startup launches new AI model - China Daily
