Previously, SiliconFlow's cloud service platform, SiliconCloud, had unveiled the full-version DeepSeek R1/V3 models based on Ascend's technology, making a substantial stride in integrating indigenous chip deployments for the DeepSeek models.

Looking back to the Spring Festival holiday on February 1, Huawei Cloud jointly announced with SiliconFlow the launch of an inference service for DeepSeek R1/V3 based on Ascend cloud servicesAround this time, the official accounts of both companies announced this news just one minute apart, emphasizing the term "first release" in their titles and summarizing the team's hard work in the content.

First Release

Yuan Jinhui, the founder of SiliconFlow, summarized that the core of SiliconFlow's technology is to provide an inference engine primarily offering high-performance LLM (Large Language Model) inference and training solutions, helping businesses effectively deploy AI applicationsDeepSeek V3, a groundbreaking open-source inference model, shocked the global tech community, and SiliconFlow's tailored services for it are especially fitting.

Before the launch of the DeepSeek V3 model, Liang Wenfeng, the founder of DeepSeek, had inquired whether SiliconFlow intended to deploy the modelAt that time, Liang even suggested having at least 20 H800 (NVIDIA GPU chips) for deployment, ideally around 80, while 10 would suffice, albeit slower.

Yuan Jinhui did some calculations: a net expenditure of five to six million yuan for 80 servers over a month seemed like a significant risk, especially since there was no guarantee they would be fully utilized, so he opted against taking the plunge.

As DeepSeek continued to create one remarkable achievement after another, Yuan found himself anxious and frustrated by the lack of sufficient computational resources until a colleague had a brainstorm: "Let's use local cards instead of relying solely on foreign hardware."

SiliconFlow proactively proposed a collaboration with Huawei Cloud.

Huawei developed the Ascend 910 and Ascend 310 AI processor chips using its self-researched Da Vinci architecture, and Huawei Cloud subsequently launched Ascend AI cloud services aimed at businesses, making AI computational power easily accessible with one-click deployment.

Huawei's Ascend AI cloud service features a “Hundreds of Models, Thousands of Scenarios” section that, aside from its proprietary Pangu model, also supports hundreds of mainstream open-source models

Advertisements

This flexibility allows enterprises and developers to create their large model applications more swiftly.

DeepSeek's immense popularity necessitated the integration with SiliconFlow, enabling both parties to synergize effectively—Huawei Cloud would allocate computational resources while SiliconFlow ensured the models would run on GPUs with robust questioning capabilities, maintaining stability and speed without compromising accuracy.

During the 2025 Spring Festival, teams from SiliconFlow and Huawei Cloud worked non-stop while the DeepSeek team provided extensive support and insights.

In the early hours of February 1, nearly ten hours before the announcement of the DeepSeek R1/V3 inference service based on Huawei Ascend, a senior executive from SiliconFlow posted on social media, stating that the platform had integrated DeepSeek series models with API service prices matching that of the DeepSeek official website.

On February 1, SiliconFlow's WeChat index skyrocketed by 8831.35%, from almost zero to peak engagementAt that time, SiliconFlow became the first platform besides DeepSeek's official channel to offer cloud services for the fully functional 671B model on domestic chips.

Riding on the wave of excitement, SiliconFlow also launched a hiring initiative with 15 full-time positions open for roles like visual generation inference engine engineers, heterogeneous hardware adaptation engineers, and R&D delivery engineers, along with eight intern positions in generative AI-related areas.

A Silent Competition in API Services

“We need to consider concurrency, along with any limitations on future concurrency.” One official from an AI application company collaborating with Huawei Cloud to integrate DeepSeek-R1 stated that these are issues every platform enterprise or application product must consider, with many underlying challenges handled primarily by Huawei Cloud.

The development of DeepSeek API services is not merely a battleground for tech giants like Tencent, Alibaba, and Baidu, as various innovative modeling firms also join in the fray.

Despite the plethora of model cloud service providers, SiliconFlow executives noted that individuals involved in model testing have begun to offer users a basis for judgment, such as whether the provided model is the original 671B parameter size

Advertisements

They also consider whether the context window size (which indicates the size of the previous token or text segment the language model processes for predictions or text generation) is appropriately tuned, and how the success rate of AI-assisted programming requests fares, including any limitations imposed by device management (typically over 100,000 devices).

Chen Tianchu, a scholar involved with large models at Zhejiang University’s ARClab, observed that different cloud providers integrating DeepSeek target various clients, with differences in cost control and pricing plansSome may provide the full version, while others could offer quantized or distilled lower-tier models.

Chen believes that beyond delivering standardized model API services, the competition among cloud providers is primarily centered on the ability to offer personalized servicesFor instance, after launching the flagship 671B DeepSeek V3/R1 model, SiliconFlow introduced six distilled versions of the DeepSeek R1 model, with the 8B, 7B, and 1.5B models offered for free, allowing businesses and developers easy API access for their AI applications.

By February 13, feedback from model testers indicated that among mainstream vendors providing DeepSeek API services, SiliconFlow’s deep collaboration with Huawei Cloud offered superior speeds in inference, generation, and response times compared to the heavily loaded official DeepSeek services, effectively creating a measurable distance from competitors like Tencent Cloud and Alibaba Cloud.

The Effects of “+DeepSeek” are Yet to be Determined

Thanks to the critical operations undertaken during the Spring Festival, Huawei launched the native HarmonyOS-based Xiaoyi Assistant App on the first workday after the holiday (February 5), integrating the inference capabilities of the DeepSeek model and featuring a Beta version of DeepSeek-R1 in its "Agents" options.

Just five days later, the pure HarmonyOS Xiaoyi App, utilizing DeepSeek-R1's capabilities, upgraded into a full version, introducing a “networked search” function that made its knowledge base far richer and more timely.

On February 13, while visiting Huawei's offline mobile phone sales store in Guangzhou, I experienced how all the updates applied to HarmonyOS-based phones enabled Xiaoyi to transform from a voice assistant service activated via the system into a standalone application

Advertisements

Accessing the Xiaoyi App now allows users to engage in conversations with "her" and directly access DeepSeek-R1 interactions without needing to download third-party applications.

Huawei utilizes its self-developed Pangu AI model to train XiaoyiWithin the pure HarmonyOS ecosystem, its interface emphasizes DeepSeek prominently after becoming an independent application.

In the reference material held by the store's sales personnel, while some comparisons were made regarding how to engage with Xiaoyi, particularly stressing Pangu's unique advantages rooted in Chinese language capabilities, no mention was made of DeepSeek's contributions.

The AI application company's official, in an interview with the Economic Observer, provided insights into how the integration of DeepSeek's model capabilities has influenced key performance indicators.

“The hallucination from the large model has decreased, and the success rate of tasks has increased,” stated the officialTheir team reported near-zero failure rates when executing tasks using DeepSeek's modelWhile acknowledging that large model hallucination remains a common problem, they have noted its reduction.

Interestingly, the AI firm, despite investing heavily in developing both general-purpose large models and specialized models for specific applications, has been shifting focus toward integrating DeepSeek's series rather than merging their existing models with itThis trend reflects a willingness to explore new application scenarios using DeepSeek.

This executive shares a consensus with Chen Tianchu that while DeepSeek is renowned for inference, its true strengths lie in being empowered by infrastructure firms like cloud service providers.

A senior platform executive who has engaged in discussions with various mainstream large model firms regarding connecting intelligent agents across different hardware terminals views this as a critical pathway for DeepSeek's future technological implementation and the subsequent growth of the current "DeepSeek" trend, “Every piece of hardware might serve as an entry point for intelligent agents, enabling dedicated AI to train tailored agents by solving human tasks.”

Why SiliconFlow?

Yuan Jinhui, the founder of SiliconFlow, is a serial entrepreneur who established OneFlow in 2017, which saw a valuation exceeding several hundred million dollars amid the 2023 boom surrounding large AI models in China.

In the same year, Wang Huiwen, co-founder of Meituan, initiated a large model company called Lightyear, suggesting that he would invite Yuan Jinhui to join as a co-founder through an acquisition of OneFlow

Advertisements

Advertisements