Yan Junjie, founder of MiniMax, gave a rare group interview with the media, clearly stating "not talking about changing the world, but discussing scientific methods." In the current environment of intense financing and DAU competition, he chose the "non-genius" team approach, gradually approaching general artificial intelligence (AGI) through cross-thinking and a full-modal technology stack. The model has now covered more than 200 countries and regions worldwide.

image.png

Cross-Thinking: Let the model reflect while running  

- Technical Highlights: Inserting a "self-reflection token" in the reasoning chain allows the model to correct logical jumps in real time  

- Industry Impact: This mechanism has been adapted by several major overseas reasoning frameworks and has become a popular plugin in the open-source community  

- Application Results: Accuracy in math competition questions increased by 12%, and hallucination rate in long document Q&A decreased by 18%

Full-Modal Approach: "Do It Right Once" Under Resource Constraints  

- Strategy: Persisting with a unified architecture for text, voice, image, and video, rejecting the shortcut of "single-modal → concatenation"  

- Verification: Voice cloning in 10 seconds and video generation at 60fps have been launched, with overseas users accounting for 65%  

- Business: Charging by usage volume, no "burning money subsidies," achieving positive cash flow when monthly active users exceed 100 million

Investor Attitude: Not chasing DAU, but pursuing "explanable AGI"  

- Funding: Latest round valuation is approximately $2.5 billion, with shareholders including Mihoyo, Tencent, and IDG  

- Pace: Funding every 18 months, "enough is enough"; not disclosing DAU, only publicizing model performance curves  

- Goal: Launch an explainable and verifiable AGI prototype by 2027, rather than "larger parameters"

Founder's Words: AGI is Science, Not Magic 

"AI is not a mysterious black box, but an engineering problem that can be analyzed using first-principle principles," said Yan Junjie. The team rejects the 'genius theory' and advocates that 'ordinary people + scientific methods' can also create extraordinary value. Currently, the MiniMax R&D team has fewer than 300 people, with all algorithms, engineering, and products developed in-house, no outsourcing.

Next Steps: Cross-Thinking 2.0 + Edge AGI