Hengyu originated from the Cave Fei Temple
Qubits | Official account QbitAI
Big models are building a new investment nugget boom.
The latest news, the large-model track startup MiniMax has completed a new round of financing of $250 million, with an overall valuation of more than $1.2 billion.
Qubit further learned that MiniMax's round of investment also successfully attracted Tencent bets, which are said to have participated with $40 million.
This is the first investment that Tencent has been exposed since the wave of large models.
At present, Tencent has made it clear that it will not miss the wave of large models, and the internal research and development of projects represented by the hybrid element assistant has been carried out, but now it seems that internal self-research is not all, Tencent once again sacrificed the secret of the mobile Internet era:
Investment and research at the same time.
Who is MiniMax?
MiniMax's current round of financing was first broken by Reuters, which quoted sources as saying that the company has completed a new round of financing with a valuation of $1.2 billion, and this round has attracted $250 million.
Earlier, Qubit also introduced that MiniMax has grown rapidly since its inception, and due to the big model heat wave driven by ChatGPT, this low-key startup is quickly becoming the most attractive star player in the venture capital field.
MiniMax takes its name from an algorithm, an AI startup founded in December 2021 called All in AGI.
More than one year after its birth, he has self-developed the basic model architecture of three modalities: text-to-visual, text-to-audio, and text-to-text, and built a computational reasoning platform on top of the basic model.
It is said that MiniMax employees first nicknamed the large model "ABAB" to imitate the early days of language training, when the model was as articulate as a baby and could only "Aba Abba".
The specific capabilities of the model were first tested on Glow, the company's to C virtual chat software product, released in November last year.
In Glow, users can create agents with backgrounds and specific personalities according to their preferences. Through content generation and user feedback, the AI capabilities of the big model behind it are constantly iterated, and there are some similar RLHF behind ChatGPT.
In the four months since its launch, the number of Glow users approached 5 million.
In February, MiniMax revealed further plans to Qubit:
This year, the API interface will be opened to the public.
In March this year, MiniMax launched an API open platform for enterprise users to support service calls for text and voice models.
Almost at the same time, in early March, another generative dialogue AI independently developed by MiniMax was launched, named Inspo, which immediately caused a lot of repercussions after the "private" closed beta.
According to the official website, Inspo's positioning is still an artificial intelligence assistant, and the large model behind it is based on Transformer.
Specific abilities include writing, getting information, giving advice, and remembering previous texts to complete continuous tasks.
After the launch of self-developed products, MiniMax also quickly disclosed a series of cooperation progress:
On April 18, Tan Zhu, president of Volcano Engine, specifically mentioned that "Volcano Engine ran Douyin and also ran MiniMax" at the "Prime Mover Conference" held by Volcano Engine;
On the same day, Kingsoft Office released WPS AI, and CEO Zhang Qingyuan also directly said that WPS AI has a mix of large models, one of which is MiniMax.
The latest news about MiniMax is that on the large-model battle platform SuperCLUE Langya list, the MiniMax model topped the national service with 1188 points, ranking second in the world.
At the top of the list was Anthropic's Claude, followed by MiniMax with a GPT of 3.5 with 1171 points.
This is the first time that a large domestic model has exceeded GPT3.5 in public evaluation.
Who is in the company?
Tencent's first investment in the large model team, what is the origin?
The core member and co-founder is Yan Junjie, former vice president of SenseTime Technology and head of general intelligent technology, and also served as vice president of SenseTime Research Institute.
Another core member of the early disclosure was Yang Bin, an alumnus of Yan Junjie at the Chinese Academy of Sciences.
In 2014, Yang Bin participated in deep learning related projects at the Institute of Automation, Chinese Academy of Sciences, and in 2016, he went to the University of Toronto to pursue a Ph.D. degree.
During his study abroad, he worked as a founding member in teams such as Uber ATG Research and autonomous driving startup Waabi.
The company registered by Inspo and Glow is Beijing Xiyu Technology, and the legal representative is Zhou Yucong.
The name is not much information available online, but it has appeared in the results of the ICCV 2019 LFR Challenge.
At that time, Zhou Yucong's team won the ICCV 2019 LFR Challenge iQIYI-Large track. This is a challenge for lightweight face recognition, Zhou Yucong's track is iQiyi Video face recognition large-scale recognition model, the team finally won the first place with an accuracy of 0.7298.
The team consists of the Multimedia Lab of Chinese University of Hong Kong and SenseTime, and Yucong Zhou is responsible for the construction and maintenance of the entire training platform.
In early March, Qubit learned that MiniMax had quickly reached unicorn-level valuation, led by Mihayou, who had already participated in a round of investment before.
But in fact, even if there is a cash flow of $1 billion, it is a drop in the bucket in front of the big model that spends money.
Perhaps this is also the reason for this, round after round of financing, and Tencent decisively struck.
And it is said that this is only the beginning of Tencent's large model investment layout, and Tsinghua's AI companies have recently appeared in rumors.
Reference Links:
[1]https://www.reuters.com/technology/china-ai-startup-minimax-raising-over-250-mln-tencent-backed-entity-others-2023-06-01/
[2]https://mp.weixin.qq.com/s/n7J5Tth2zmuOREtwugLbwg
— End —
Qubits QbitAI · Headline number signed
Follow us and be the first to know the latest scientific and technological trends