In a conversation with Lawrence Burns, investment manager at Scottish Mortgage, MiniMax founder Yan Junjie said that China's leading large language models have reached roughly 95% of the performance of top U.S. models at only one-tenth the cost. However, closing the remaining 5% gap is far harder than getting the first 95% of the way there. Both Chinese and U.S. large models remain in the early stages on the road to AGI, and their current pace of progress is broadly comparable. Constrained by computing power, MiniMax puts efficiency at the heart of everything it does: it boosts compute utilization by optimizing model architecture and building its own infrastructure, and reinvests the inference costs it saves into post-training, with the goal of building a post-training system of leading scale. Its team of just 500 people also reflects a relentless focus on organizational efficiency. MiniMax's annualized revenue run rate now exceeds US$800 million, with enterprise business contributing more than 80%. Yan said the company will cut costs through technological innovation and grow revenue faster than costs rise, though he has not yet set a timetable for profitability. He traced his original motivation for founding the company to his grandfather, who wanted to write a memoir but had no AI technology at the time that could help him. After the company was founded, it began by pursuing multimodal technology on several fronts in parallel. Its current models fall into two lines: the M series, focused on coding and agents, and the H series, centered mainly on video generation. The company is now developing its next-generation model, M3 Pro. In Yan's view, the test of AGI is whether AI can produce new theories and become the primary foundation for new scientific breakthroughs. MiniMax is committed to continuous open sourcing until AGI is achieved. Open sourcing not only draws a large number of developers into model iteration and gives rise to a rich variety of application scenarios, but also meets the data privacy and controllability needs of government and enterprise customers. Yan believes Chinese and U.S. AI will take different paths: the U.S. ecosystem relies on massive capital to build closed systems, while the Chinese ecosystem relies on efficiency to build a collaborative one. AGI technology, he argues, should not be monopolized by any single entity, and the fruits of intelligent development should be shared by all.