On the eve of DeepSeek triggering domestic adaptation, "Silicon-Based Flow" has completed a financing of 100 million yuan | First Release by Zhiyong.

Written by Deng Yongyi

Edited by Su Jianxun

Did anyone expect DeepSeek to gain global popularity? At least, before the release of DeepSeek-V3, Yuan Jinhui, who was "selling shovels" on the front line of the large model battlefield, was also uncertain.

On January 28, he recalled a story on social media: One month before DeepSeek became popular, Liang Wenfeng suggested that SiliconFlow deploy DeepSeek-V3, and prepare at least 20 H800s, and 80 would be the best. Yuan Jinhui quickly calculated that 80 H800 servers in one month would cost a computing power cost of five to six million yuan, and these machines might not be fully utilized. "The risk is too great, so I didn't take the bet."

Then DeepSeek became popular worldwide, and the surging traffic caused its official website and App to crash quickly. Enthusiastic users were eager to find a channel where DeepSeek could be used normally. Yuan Jinhui lamented: "Now DeepSeek is so popular, it's a decision-making error, and I'm heartbroken."

Soon, this remorse turned into a powerful action. SiliconFlow, which was suffering from insufficient computing power, found Huawei and fully completed the adaptation work of DeepSeek-R1 & V3 on the Ascend chip. During the entire Spring Festival, the SiliconFlow team did not rest for a day.

Their actions were fast enough to catch the first wave of traffic. On February 1, SiliconFlow released the full-powered version of DeepSeek-R1 & V3 based on Huawei's Ascend computing power, which quickly attracted a large number of developers and individual users.

△ Source: SimilarWeb

This is a smart cooperation that once and for all clears the market's doubts about the performance bottleneck of domestic chips, proving that deploying DeepSeek on domestic chips can also achieve the effect of being on par with the global high-end GPU deployment model. At the same time, this cooperation accidentally triggered a carnival of domestic substitution.

SiliconFlow also used this to prove its AI Infra strength. "Especially after the launch of DeepSeek R1 & V3 this time, the cooperation leads that come in are temporarily too many to handle." Yuan Jinhui said.

Before SiliconFlow became popular, "Intelligent Emergence" learned that SiliconFlow had completed a Pre-A round of financing of 100 million yuan in RMB by the end of 2024, led by Hua Chuang Capital, followed by Pu Hua Capital, and the old shareholder Yao Tu Capital continued to oversubscribe. Hua Xing Capital served as the exclusive financial advisor.

It is worth noting that before this round of financing, SiliconFlow had introduced Meituan as a strategic shareholder.

In August 2024, "Intelligent Emergence" once interviewed the founder Yuan Jinhui. At that time, SiliconFlow had just celebrated its first anniversary and just released the MaaS platform SiliconCloud.

You can imagine SiliconCloud as a "big model API supermarket". Developers only need to register an account, choose the model they want, and they can develop applications, eliminating the burden of buying a server. Even the models adapted and optimized by SiliconFlow can respond faster and have a lower reasoning cost than the original factory models.

In this year, SiliconFlow has experienced at least a tenfold market growth. SiliconCloud has gained more than three million users, with an average daily call of more than 100 billion Tokens - when it was just launched last year, this number was still an average daily of 10 billion.

△ Source: SiliconFlow

In the developer community, SiliconFlow has a good reputation and is called the "Cyber Bodhisattva" for providing model services with affordable prices and excellent performance. Now, SiliconCloud has launched hundreds of mainstream models including the full-powered version of DeepSeek-R1 & V3, Qwen, FLUX.1, CosyVoice, etc., and also provides a full set of tools for model training, fine-tuning, hosting to deployment to meet the different application scenario needs of developers and enterprise customers.

In July 2024, SiliconFlow also launched a new product: the ComfyUI-based cloud node plugin BizyAir.

BizyAir is an image generation tool that enables AI developers and creators to directly use cloud computing resources to complete image generation tasks without considering graphics card and environmental limitations. Up to now, BizyAir supports dozens of models, supports the local node mixed running of ComfyUI, and has gained nearly 10,000 users.

△ Source: SiliconFlow

In this year, SiliconFlow has also started the To B commercial expansion and established partnerships with top cloud vendors, chip vendors, large model companies and application companies such as Huawei, Amazon Web Services, NVIDIA, etc.

In terms of enterprise services, SiliconFlow has landed benchmark customer cases in multiple scenarios such as large language models, text-to-image, and video generation. SiliconFlow currently provides four service models, including exclusive instances, computing power management, private version MaaS, and an all-in-one machine based on Huawei's Ascend 910 series NPU.

There is no doubt that DeepSeek has swept away the haze in the large model field in 2024, and the prelude to a grand journey has begun. In 2024, Yuan Jinhui once told "Intelligent Emergence" that after the emergence of GPT-4o, the limitations of the language ability and knowledge ability of large models have been better solved, except for the logical reasoning ability.

The emergence of OpenAI o1 and DeepSeek R1 means that large models have taken a big step forward in the "logical ability" issue, which is a great benefit for the application market.