MiniMax's Yan Junjie Says China's Last 5% AI Gap Is the Hardest
Yan Junjie, founder of the Chinese AI startup MiniMax, has a blunt way of describing where his country’s large language models stand against America’s best. In a recent episode of the British podcast Invest in Progress, he told Scottish Mortgage investment manager Lawrence Burns that China’s leading models have already closed roughly 95 percent of the distance to the top US systems. The remaining 5 percent, he argued, may prove harder to close than everything that came before it.
The conversation was recorded after Burns traveled to China in December 2025 and met Yan in person. What stuck with the investor was not a product demo but a number: the estimate that the gap between the two countries’ best models is about 5 percent. Yan’s framing turns that small figure into the central problem of Chinese AI rather than a victory lap. The last stretch, in his telling, is where the hardest engineering and the deepest uncertainty sit.
That view fits the wider mood in China’s AI industry, where companies have spent the past two years releasing models at a furious pace and claiming benchmark parity with US rivals, only to find that the final increment of quality is the most expensive. MiniMax itself has lived through that cycle. The company has shipped a series of models aimed at both consumer apps and developer tools, and it has been public about the pressure to keep pace with far better-funded competitors.
The 5 percent figure also matters because it reframes how the AI race is discussed. If China’s best models are already within a few points of the frontier, then the contest is no longer about catching up on raw capability but about reliability, cost, and the unglamorous work of making models trustworthy in real deployments. Yan’s point is that this last phase rewards different strengths than the scramble that preceded it.
MiniMax’s position is not that of a giant. It operates alongside larger Chinese players and faces the same constraints on advanced chips that shape the whole domestic industry. Yet it has kept a high public profile, and its founder’s willingness to put a precise number on the gap — and to say the hardest part is still ahead — is unusual in a sector where executives often prefer to talk about breakthroughs.
For global readers, the interview offers a rare candid assessment from inside China’s model-building community. Burns, whose firm invests globally, came away with the impression that the gap is real but narrow. The question Yan’s argument leaves open is what closing that final margin actually requires: more compute, better data, smarter training methods, or simply time.
Yan did not offer a timeline for when the last 5 percent might be closed, and his assessment is a judgment rather than a measured result. But the framing is likely to travel. As Chinese labs release their next generation of large language models, the industry will be watching whether the distance to the frontier keeps shrinking or whether, as Yan suggests, the final stretch proves to be the longest.