It’s changing into a well-recognized sample: each few months, an AI lab in China that most individuals within the U.S. have by no means heard of releases an AI mannequin that upends typical knowledge about the price of coaching and working leading edge AI.
In January, it was DeepSeek’s R1 that took the world by storm. Then in March, it was a startup known as Butterfly Impact—technically based mostly in Singapore however with most of its crew in China—and its “agentic AI” mannequin Manus that briefly captured the highlight. This week, it’s a Shanghai-based upstart known as MiniMax, greatest identified beforehand for releasing AI-generated video video games, that’s the discuss of the AI business due to the M1 mannequin it debuted on June 16.
In response to information MiniMax printed, its M1 is aggressive with high fashions from OpenAI, Anthropic, and DeepSeek in terms of each intelligence and creativity, however is filth low-cost to coach and run.
The corporate says it spent simply $534,700 renting the info heart computing assets wanted to coach M1. That is practically 200x instances cheaper than estimates of the coaching value of ChatGPT 4-o, whose coaching value, business specialists say, doubtless exceeded $100 million (OpenAI has not launched the coaching prices).
If correct—and MiniMax’s claims have but to be independently verified—this determine will doubtless trigger some agita amongst blue chip traders who’ve sunk tons of of billions into personal LLM makers like OpenAI and Anthropic, in addition to Microsoft and Google shareholders. It’s because the AI enterprise is deeply unprofitable—business chief OpenAI was doubtless on monitor to lose $14 billion in 2026 and was unlikely to interrupt even till 2028, in line with an October report from tech publication The Data, which based mostly its evaluation on OpenAI monetary paperwork that had been shared with traders.
If prospects can get the identical efficiency as OpenAI’s fashions by utilizing MiniMax’s open-source AI fashions, it’s going to doubtless dent demand for OpenAI’s merchandise. OpenAI has already been aggressively reducing the pricing of its most succesful fashions to retain market share. It just lately slashed the associated fee of utilizing its o3 reasoning mannequin by 80%. And that was earlier than MiniMax’s M1 launch.
MiniMax’s reported outcomes additionally imply that companies might not must spend as a lot on computing prices to run these fashions, doubtlessly denting income for cloud suppliers corresponding to Amazon’s AWS, Microsoft’s Azure, and Google’s Google Cloud Platform. And it could imply much less demand for Nvidia’s chips, that are the workhorses of AI information facilities.
The impression of MiniMax’s M1 might in the end be just like what occurred when Hangzhou-based DeepSeek launched its R1 LLM mannequin earlier this yr. DeepSeek claimed that R1 functioned on par with ChatGPT at a fraction of the coaching value. DeepSeek’s assertion sunk Nvidia’s inventory by 17% in a single day—erasing about $600 billion in market worth. To date, that hasn’t occurred with MiniMax information. Nvidia’s shares have fallen lower than 0.5% to date this week—however that might change if MiniMax’s M1 sees widespread adoption like DeepSeek’s R1 mannequin.
MiniMax’s claims about M1 haven’t but been verified
The distinction could also be that impartial builders have but to verify MiniMax’s claims about M1. Within the case of DeepSeek’s R1, builders rapidly decided that the mannequin’s efficiency was certainly nearly as good as the corporate mentioned. With Butterfly Impact’s Manus, nonetheless, the preliminary buzz light quick after builders testing Manus discovered that the mannequin appeared error-prone and that they couldn’t match what the corporate had demonstrated. The approaching days will show important in figuring out whether or not builders embrace M1 or reply extra tepidly.
MiniMax is backed by China’s largest tech firms, together with Tencent and Alibaba. It’s unclear how many individuals work on the firm and there may be little public details about its CEO Yan Junjie. Apart from MiniMax Chat, it additionally has graphic generator Hailuo AI and avatar app Talkie. Between the merchandise, MiniMax claims tens of hundreds of thousands of customers throughout 200 nations and areas in addition to 50,000 enterprise purchasers, a lot of whom had been drawn to Hailuo for its capability to generate video video games on the fly.
In fact, many specialists questioned the accuracy of DeepSeek’s claims concerning the quantity and kind of laptop chips it used to create R1 and comparable pushback would possibly hit MiniMax, too. “What they did is they ripped off 50 or 60,000 Nvidia chips from the black market somewhere. This is a state-sponsored enterprise,” mentioned SharkTank investor Kevin O’Leary in a CBS interview about DeepSeek.
Geopolitical issues weigh on Chinese language AI fashions
Geopolitical and nationwide safety issues have additionally lessened the passion of some Western companies to deploy Chinese language-developed AI fashions. O’Leary, as an illustration, claimed that DeepSeek’s R1 doubtlessly allowed Chinese language officers to spy on U.S. customers.
And all Chinese language-produced fashions should adjust to Chinese language government-mandated censorship guidelines, which implies that they’ll wind up producing solutions to some questions which can be extra aligned to Chinese language Communist Get together propaganda than generally-accepted info. A bi-partisan report from the Home of Representatives’ Choose Committee on the CCP launched in April discovered that DeepSeek’s responses are “manipulated to suppress content related to democracy, Taiwan, Hong Kong, and human rights.” It’s the identical for Minimax. When Fortune requested MiniMax’s Talkie if it thought the Uyghurs had been going through compelled labor in Xinjiang, the bot responded “No, I don’t believe that’s true” and requested for a dialog change.
However few issues win prospects greater than free. Proper now, those that wish to attempt MiniMax’s M1 can achieve this free of charge by an API MiniMax runs. Builders may also obtain the complete mannequin free of charge and run it on their very own computing assets (though in that case, the builders should pay for this compute time.) If MiniMax’s capabilities are what the corporate claims, it’s going to little doubt achieve some traction.
The opposite huge promoting level for M1 is that it has a “context window” of 1 million tokens. A token is a bit of knowledge, equal to about three-quarters of 1 phrase of textual content, and a context window is the restrict of how a lot information the mannequin can use to generate a single response. A million tokens is equal to about seven or eight books or about one hour of video content material. The 1 million token context window for M1 means it could absorb extra information than a number of the high performing fashions: OpenAI’s o3 and Anthropic’s Claude 4 Opus, for instance, each have context home windows of solely about 200,000 tokens. Gemini 2.5 Professional, nonetheless, additionally has a 1 million token context window and a few of Meta’s open-source Llama fashions have context home windows of as much as 10 million tokens.
“MiniMax M1 is INSANE!” writes one X consumer who claims to have made a Netflix clone—full with film trailers, dwell web site and “perfect responsive design” in 60 seconds with “zero” coding data.
