The founder of one of China’s four leading model companies has restated the gap to the American frontier in a way that is less flattering than the headline comparison suggests.
The five per cent problem
Yan Junjie, founder of MiniMax, spoke on the UK podcast Invest in Progress with Lawrence Burns, an investment manager at Scottish Mortgage.
Burns had visited China in December 2025 and came away with a specific judgement from Yan: the best Chinese models were reaching about 95 per cent of the performance of the top American models at roughly one tenth of the cost. Scottish Mortgage subsequently participated in MiniMax’s Hong Kong listing.
On the return conversation, Burns asked how those two gaps had moved. Yan did not offer a new percentage. He said the final 5 per cent still requires time, and that both Chinese and American models remain early on the path to AGI. What he watches is the rate of progress rather than the distance.
Compute as the binding constraint
Yan returned repeatedly to compute. He described it as a permanent constraint for a Chinese AI company, and said MiniMax has designed from its founding around extracting more work from a fixed amount of it.
One design principle is that a model should perform well regardless of which chip it runs on. Chinese model companies have each chosen a different route to efficiency, some adjusting transformer architecture, some rebuilding post-training infrastructure, some concentrating on particular modalities. That divergence is why the Chinese models still differ from one another.
Burns put the question directly, asking whether American frontier labs have been spoiled by abundant compute and never forced to make efficiency central. Yan kept the answer on his own position, saying that harder hardware constraints make better engineers and researchers.
Cheap inference as an input, not an outcome
The cost reduction is not the end goal. Yan wants the savings in inference to be spent on post-training, because reinforcement learning and synthetic data generation both consume inference capacity. A lower cost per call means more attempts per unit of compute.
He expects post-training to carry a growing share of model improvement, and raised the possibility that MiniMax could eventually run the largest post-training system in the industry.
The same logic shapes the organisation. MiniMax has about 500 staff. Yan described two rules: keep architecture and infrastructure as simple as possible, and build infrastructure in-house even when that means slower product releases.
The numbers behind it
As of August, annualised revenue run rate exceeded USD 800 million, with enterprise business contributing more than 80 per cent. The company notes this figure annualises August revenue and does not represent full-year revenue of that size.
Disclosed first half 2026 revenue was USD 116.6 million, up 283.1 per cent year on year. Open platform and other AI enterprise services contributed USD 73.9 million, or 63.4 per cent of the half. Adjusted net loss for the period was USD 293 million.
Yan did not offer a date for profitability. His stated route is to lower inference cost through technical work, hold a price advantage while improving capability, and let revenue grow faster than cost.
Open source as a development method
MiniMax charges for API access while releasing model weights. Yan argues that opening the weights supports commercial adoption and lets developers explore work the company cannot staff itself.
The H3 model was downloaded about 25 million times in the month before the interview, and around 300 derivative open models were built on it within the first two weeks. One example is H3 Max, produced by the American company fal.ai through post-training, which generates five seconds of video in three seconds, faster than real time.
That speed opens interactive uses. Developers have already built 24-hour AI streams that generate footage in response to what viewers ask for.
Yan committed to continuing open releases: the company intends to keep open-sourcing until AGI is reached.
The strategic reading is narrower than the slogan. For a company constrained on compute, giving away weights converts outside developers into free capability discovery, and their findings return as direction for the next training run.
Editor’s note: This is an adapted translation of the original 36Kr report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://36kr.com/p/4060696576038533.