Wang Huiwen took a 53-day sick leave, and his OneFlow team started a new business

Source: Qubit

The core team OneFlow acquired by Wang Huiwen light years away will start a new business in the near future.

The news came from Yuan Jinhui, the founder of Oneflow, and was forwarded immediately by Yao Chi, a R&D engineer in the team.

Then Yuan Jinhui also responded immediately, emphasizing that the new entrepreneurial project is aimed at the cost of large-scale model reasoning.

OneFlow is a well-known open source deep learning framework and development platform in China, and the company behind it is first-class technology.

Previously, Wang Huiwen acquired 46.52% of the shares of First-Class Technology. The current legal representative of the company is Wang Huiwen.

After the merger and acquisition of first-class technology, there is a core technical team light years away.

The progress of financing has also accelerated. A new round of financing was completed in early June, and ** Entrepreneurship Hundred Days was successfully promoted to unicorn. **

On June 26, just when people were expecting that funds and talents from light years away would be in place and ready to show their strengths, the news of the founder Wang Huiwen’s illness and retirement came out.

Three days later, Meituan acquired 100% of Light Years Beyond for RMB 2.058 billion, and the first-class technology OneFlow team as its core asset was also transferred to Meituan.

Another 50 days later, Yuan Jinhui announced that he would lead the original team of OneFlow to start a business again. It is unknown whether he will leave Meituan or establish a new company.

In addition, Qichacha shows that the address of the first-class technology company has also changed recently, moving from Wudaokou to Zhongguancun next door.

OneFlow started a new business, aiming at AI reasoning

Yuan Jinhui, the founder and CEO of First-Class Technology, is a well-known AI architecture expert (screen name "Teacher Wood"). He graduated with a Ph.D. from the Department of Computer Science and Technology of Tsinghua University, and studied under Academician Zhang Bo, the founder of artificial intelligence in China.

Before founding First-rate Technology, he was a researcher in charge of Microsoft Research Asia, during which he developed LightLDA, the fastest topic model training algorithm at that time, and built a deep learning platform based on heterogeneous clusters, which won the special award from the president of Microsoft Research Asia.

Yuan Jinhui was also responsible for the research and development of the Chinese Eagle Eye system, which is widely used in sports events and the daily training of the national team.

Since its establishment in 2017, First-Rate Technology has launched OneFlow, a heterogeneous distributed deep learning framework, OneBrain, an MLOps full lifecycle management platform, and Libai, an open source toolbox for large-scale model training.

In short, it is more focused on the training stage of AI.

When Oneflow first started its business, the industry did not have a clear concept of "big model". At that time, the visual model based on the convolutional neural network architecture was the mainstay, and the parameter scale was tens of millions, even if it was large. Efficient training of larger models is the direction Yuan Jinhui is optimistic about. .

Now that the team is starting a business for the second time, domestic and foreign open-source and closed-source large models have been trained and blossomed everywhere, and it is a natural choice to turn to the reasoning and deployment of AI applications.

AI Infrastructure Entrepreneurship is Hot

Regardless of AI training and reasoning, it can be attributed to a larger concept, AI Infrastructure (AI Infra).

Today, in the large-model entrepreneurship market, in addition to research and development of basic models and development of large-scale model applications, large-scale infrastructure entrepreneurship is also ushering in a new wave of explosion.

For example, Lepton AI, a start-up company founded by Jia Yangqing, former VP of Ali, is entering AI Infra.

Similar to Yuan Jinhui, Jia Yangqing was first known as the master of AI frameworks. During his doctoral period, he created the well-known deep learning framework Caffe, and was a core member of the mainstream frameworks TensorFlow and Pytorch when he was working at Google and Facebook.

Now, judging from the only brief introduction on Lepton AI's official website, "running artificial intelligence applications efficiently and on a large scale" is also aimed at the inference deployment of large model applications.

The open source alpaca LlaMA2 large model deployed by Lepton AI was evaluated by the CEO of Vercel as "the fastest I have ever seen".

(The picture is not accelerated, the sparrow eats quickly)

Speaking of mainstream deep learning frameworks, a group of members of the Pytorch team recently left to start the AI Infra company fireworks.ai.

Among the three joint ventures, CEO Lin Qiao was previously the senior engineering director of Meta, Dmytro Dzhulgakov was the technical director, and Ning Li was the vice president of engineering.

However, unlike OneFlow's new project and Lepton AI, according to the official website information, fireworks.ai is currently more focused on the fine-tuning of large models.

Reference link: [1] [2] [3] [4]