China’s MiniMax debuts M1 model it says cost 200x less to train than OpenAI’s GPT-4
It has become a familiar pattern: every few months, the Amnesty International Laboratory in China has never heard of most people in the United States about the issuance of the Amnesty International model that increases the traditional wisdom about the cost of training and the advanced AI.
In January, it was the R1 Deepseek that took the world by storm. Then in March, a startup called the butterfly effect – its headquarters in Singapore, but with most of its team in China – and the “AISERIC AI” model, Manus, which took the spotlight for a short time. This week, it is an envelope based in Shanghai called Minimax, which was previously known for the release of video games created from artificial intelligence, and this is talking about the artificial intelligence industry thanks to the M1 model that first appeared on June 16.
According to the data published by the Minimax, the M1 has a competitive with the best models of Openai, Noterops, and Deepseek when it comes to intelligence and creativity, but it is cheap for training and operation.
The company says it has spent only 534,700 dollars in renting computing resources at the Data Center needed to train the M1. This is approximately 200 times cheaper than the cost of ChatGPT-4O training, which, as industry experts say, is likely to exceed $ 100 million (Openai has not issued the cost of training costs).
If the accuracy and demands of Minimax are not yet verified independently-it is likely that this number in some investors of the blue chip who drowned hundreds of billions to LLM makers such as Openai and Anthropor, as well as Microsoft and Google shareholders. This is because artificial intelligence works are profitable in depth; The Openai leader is likely to lose $ 14 billion in 2026, and it is unlikely to be broken until 2028, according to a report issued by Stocer Tech SubTation Information, which depends on its analysis on the financial documents that were shared with investors.
If customers can get the same performance as Openai models using open source AI models in Minimax, it is possible that you require demand for Openai products. Openai has already reduced the prices of its most able models to keep the market share. It has recently reduced the cost of using the O3 thinking model by 80 %. This was before the launch of M1 Minax.
Minimax results also mean that companies may not need to spend a lot on computing costs to operate these models, and may have possible profits for cloud services providers such as Amazon’s Aws and Microsoft’s Azure and Google Cloud. This may mean less on NVIDIA chips, which are the work of artificial intelligence databases.
The MINAX M1 effect may eventually be similar to what happened when the Hangzhou -based Deepseek has released the R1 LLM model earlier this year. Deepseek claimed that R1 works equally with Chatgpt in a small part of the training cost. The Dibsic statement sank 17 % NVIDIA shares in one day, achieving about $ 600 billion in the market value. So far, this has not happened with Minimax news. NVIDIA shares have decreased from less than 0.5 % so far this week – but this may change if the M1 Miminerax sees widespread dependence like the R1’s Deepseek model.
Minimax claims have not yet been verified
The difference may be that independent developers have not yet confirmed Miniveax’s claims about M1. In the case of Deepseek’s R1, the developers quickly decided that the model’s performance was already good as the company said. With Manus Butterfly Effect, the primary ton faded quickly after the developers found the Manus test that the model seemed to be at risk of error and was unable to match what the company showed. The coming days will prove crucial in determining whether the developers are hugging M1 or more wondering.
Minimax supports the largest technical companies in China, including TENCENT and Alibaba. The number of people working in the company is not clear, and there is a few general information about its CEO, Jan Jonji. Regardless of the Minimax chat, the company also offers the Hailuo Ai and Avatar App Talkie. Through these products, Minimax claims that tens of millions of users in 200 countries and regions in addition to 50,000 customers from the institution, a number of them were withdrawn to Hailuo because of their ability to generate video games while flying.
Of course, many experts questioned the accuracy of Deepseek’s claims about the amount and type of computer chips that they used to create R1, and a similar reaction may reach a minimum. “What they did is that they separated from 50 or 60,000 NVIDIA chips from the black market somewhere. This is a country sponsored by the state,” he said. shark Investor Kevin Olieri in an interview with CBS about Deepseek.
Geopolitical considerations weigh the Chinese artificial intelligence models
The concerns of geopolitical and national security also reduced the enthusiasm of some Western companies to spread artificial intelligence models developed by Chinese. For example, O’Leary claimed that the R1 Deepseek allowed Chinese officials to spy on American users.
All models produced by the Chinese must comply with the Chinese control rules imposed by the Chinese government, which means that they can end with answers to some questions that are compatible with Chinese communist party propaganda more than the facts accepted in general. A report from the two parties from a selected committee from the House of Representatives on CCP in April found that Dibsic’s responses “are tampered with to suppress the content related to democracy, Taiwan, Hong Kong and human rights.” It is the same for Minimax. when luck When asked about the talk about Minimax if he believed that the Uighurs face forced work in Xinjiang, the robot replied, “No, I don’t think this is true” and asked to change a conversation.
But a few things win customers more than free access. Currently, those who want to try MIIX M1 from Mimimorax can do this for free by running API Minimax. The developers can also download the entire form for free and operate it on their computing resources (although in this case, developers have to pay for the calculation time). If Minimax’s potential is what the company claims, it will undoubtedly get some traction.
Another large M1 sale point is that it contains a “context window” of a million icons. The distinctive symbol is a large part of the data, which is equivalent to about three quarters of one word from the text, and the window of context is to reduce the amount of data that the model can use to create one response. One million symbols equivalent to about seven or eight books or one hour of video content. The context window, which is one million people for M1, means that it can take more data than some of the higher performance models: Openai’s O3 and Claude OPUS 4, for example, both contains context windows of about 200,000 symbols. However, Gemini 2.5 Pro also has a context window of one million million, and some Llama open source models in Meta have context windows of up to 10 million icons.
“Minimum M1 is crazy!” One of the X -user uses he writes that he has created Netflix cloning – with movie trailers, direct web site, and “ideal, responsive design” in 60 seconds with “Zero” with coding.
2025-06-18 16:56:00
