Author | Liu Baodan
Editor | Huang Yu
46 days ago, DeepSeek-R1 was officially released, and the model weights were simultaneously open-sourced, with the app also updated. Since then, DeepSeek has gained global popularity, leading Chinese AI to take a significant step forward. 46 days later, Alibaba replicated this path.
In the early hours of March 6, Alibaba released and open-sourced a new inference model, Tongyi Qianwen QwQ-32B. This model's overall performance is on par with DeepSeek-R1, and the deployment cost has been further reduced to consumer-grade graphics card levels. At the same time, users will be able to experience this model for free through the Tongyi app.
This means that after DeepSeek, Tencent, and other companies like Moon's Dark Side, Alibaba has officially launched a deep inference model, further promoting the application of models in more complex scenarios.
Qianwen QwQ-32B is Alibaba's latest achievement in exploring inference models, with inference capabilities combined with low consumption, making this model suitable for applications that require quick responses or high data security. However, several industry insiders have expressed to Wall Street Insights that the market feedback for this model will need some time to observe.
At the close of trading that day, Alibaba's Hong Kong stock was HKD 140.800, reaching a nearly three-year high, with a stock price increase of 8.39% for the day. Since the beginning of the year, Alibaba's stock price has risen over 70%, with a market capitalization of HKD 2.68 trillion. In the U.S. stock market close, Alibaba's stock price slightly fell by 0.77%.
This global AI competition is currently at a critical stage of breakthrough, and Alibaba's tough battle has just begun.
<h2>Catching Up with DeepSeek</h2>
DeepSeek is currently the most favored AI large model, and now, Alibaba is preparing to take its place.
This time, Alibaba's Tongyi Qianwen QwQ-32B belongs to the medium-parameter model category, and its biggest highlight is achieving effects that only large-parameter inference models possess, largely proving that parameter scale is no longer a decisive factor for model performance.
According to a series of authoritative benchmark tests released by Alibaba, the Qianwen QwQ-32B model performed exceptionally well, almost completely surpassing OpenAI-o1-mini, and is on par with the strongest open-source inference model DeepSeek-R1, which has 671 billion parameters.
For example, in the AIME24 evaluation set testing mathematical abilities and the LiveCodeBench assessing coding abilities, Qianwen QwQ-32B performed comparably to DeepSeek-R1, far surpassing o1-mini and the same-sized R1 distilled model. In the LiveBench, Google's proposed IFEval evaluation set, and the BFCL test proposed by the University of California, Berkeley, Qianwen QwQ-32B's scores exceeded those of DeepSeek-R1.
While maintaining strong performance, Qianwen QwQ-32B has significantly reduced deployment and usage costs. Alibaba stated that this model can also achieve local deployment on consumer-grade graphics cards
Liu Daoru, founder and CEO of Beijing Xinghan Future, told Wall Street Insights that the DeepSeek-R1 full version is 671B, which has very high deployment configuration requirements, needing at least 8 A100 graphics cards. For QwQ-32B, an enhanced version of the NVIDIA 4090 can be deployed, costing less than 1/10 of R1, which will be more conducive to the application and popularization of inference models.
The higher cost-performance ratio of QwQ-32B is attributed to Alibaba adopting a different technical route. Wall Street Insights learned from insiders that this model uses a dense architecture, while DeepSeek is a mixture of experts system (MoE). Dense and MoE are two forms of model architecture.
The insider further stated that the Alibaba Tongyi team, based on cold start, conducted two rounds of large-scale reinforcement learning targeting mathematical and programming tasks and general capabilities, achieving surprising improvements in inference capabilities at the 32B model size, confirming that large-scale reinforcement learning can significantly enhance model performance.
An internal employee of Alibaba Cloud mentioned that the inference model has been in development for a long time, with the company starting overtime work since the second day of the Lunar New Year, and has not finished work before 2 AM since then. The company released a preview version last week, followed by the official version and open-source release this week.
In the industry’s view, QwQ-32B further reduces the application cost of models, especially as this model can provide strong inference capabilities while meeting lower resource consumption requirements, making it suitable for applications that require quick responses or have high data security demands. Developers and enterprises can deploy it on consumer-grade hardware to create AI solutions.
Currently, QwQ-32B has been open-sourced on platforms such as Modao Community, HuggingFace, and GitHub under the permissive Apache 2.0 license, allowing everyone to download the model for local deployment for free. Users will also be able to experience the latest QwQ-32B model for free through the Tongyi APP.
For cloud deployment needs, users can quickly deploy through the Alibaba Cloud PAI platform and perform model fine-tuning, evaluation, and application building; or choose container service ACK paired with Alibaba Cloud GPU computing power to achieve model containerization deployment and efficient inference.
<h2>Three Years of Spending 380 Billion</h2>
If DeepSeek has ignited the Chinese AI industry, then Alibaba is undoubtedly one of the winners reaping the benefits of this AI boom.
On February 24, Alibaba Group CEO Eddie Wu announced that in the next three years, Alibaba will invest over 380 billion yuan to build cloud and AI hardware infrastructure, exceeding the total amount of the past decade. This also sets a record for the largest investment ever made by a private enterprise in China in the field of cloud and AI hardware infrastructure construction.
Alibaba has made extensive layouts in AI, including cloud computing, models, and applications within its ecosystem, but ultimately, the main line that Alibaba values most is the cloud computing market that supports AI applications
On February 21, during the latest quarterly conference call, Alibaba stated that the primary goal of its AI strategy is to pursue the realization of AGI, continuously breaking through the boundaries of model intelligence capabilities. Intelligence refers to the tokens output by AI models, and in the future, 90% of tokens will be generated and output on cloud computing networks. Only through Alibaba Cloud's globally distributed data centers can these be delivered to global developers more quickly.
For Alibaba, an important strategy is open-source. As one of the earliest large companies in China to open-source self-developed large models, Alibaba Cloud has taken the lead in the industry to achieve &#34;full-size, multi-modal, and multi-scenario&#34; open-source.
Before QwQ-32B, Alibaba had already open-sourced the Qwen2.5-1M model with a context of about 1 million tokens, as well as the visual model Qwen2.5-VL this year. In November last year, Alibaba open-sourced the entire series of Tongyi Qianwen code models, totaling 6 Qwen2.5-Coder models.
As of now in 2023, the Tongyi team at Alibaba has open-sourced over 200 models, including the large language model Qwen and the visual generation model Wan, covering two major foundational model series. The open-source includes text generation models, visual understanding/generation models, speech understanding/generation models, text-to-image and video models, etc., with parameters ranging from 0.5B to 110B.
According to insiders at Alibaba, as of now, the number of derivative models of Qwen in the domestic and international AI open-source community has exceeded 100,000, surpassing the Llama series models from the United States, becoming the largest open-source model group in the world.
Liu Daoru believes that the most beneficial aspect of model open-sourcing is for cloud vendors like Alibaba Cloud. While the threshold for models has been lowered, the threshold for computing power remains very high. Both inference and training require large-scale GPU computing power, which is why Alibaba is fully promoting open-source.
Alibaba Cloud's Chief Technology Officer Zhou Jingren once told Wall Street Journal that he hopes to open-source advanced technologies with an open mindset, allowing everyone to explore in parallel. The same goes for AI products; through an open-source system, they can explore with enterprises, focusing not only on improving model capabilities but also on what can be done based on the models, deeply exploring their prospects and potential.
However, the effects of Alibaba's QwQ-32B after open-sourcing still need time for further observation. Industry insiders have expressed concerns that current inference models are mainly used for solving mathematical problems and writing code, but what DeepSeek has excelled in is literary creation and ideological content, making it not easy to achieve a &#34;peer&#34; effect from the user's perspective.
What is certain is that open-source has become the mainstream of AI development. DeepSeek announced its latest technological progress through an open-source week event, including four open-source projects: FlashMLA, DeepEP, DeepGEMM, and 3FS, as well as code repositories like DualPipe and EPLB, and disclosed a theoretical cost-profit margin of 545% on the last day
On March 6, Tencent's Hunyuan released the image-to-video model and made it open source. Currently, the Hunyuan open-source series models have accumulated over 23,000 developer followers and stars on GitHub. Baidu recently announced that the Wenxin large model 4.5 will be released on March 16 and will officially be open-sourced on June 30.
On February 18, StepStar first opened source, releasing the open-source video generation model Step-Video-T2V and the open-source voice interaction large model Step-Audio to global developers. Earlier, MiniMax open-sourced the new MiniMax-01 series models, including the foundational language large model MiniMax-Text-01 and the visual multimodal large model MiniMax-VL-01.
At the same time, the Dark Side of the Moon also announced the sparse attention technology—MoBA (Mixture of Block Attention), which is an attention architecture inspired by Mixture of Experts (MoE) and Block Sparse Attention, capable of seamlessly switching between full attention and sparse attention modes, making it an effective solution for long-context tasks.
As various AI companies unveil their underlying technological routes, the AI industry has made significant strides from the initial technological competition to practical applications, which will be the most intense battleground for AI companies by 2025.
<h2>The Battle for AI Applications Begins</h2>
Open source has become the choice for most AI enterprises, and for many, whether to open source is seen as crucial for a company to take the lead in this AI arms race.
According to insiders, the industry's early move to open source was a reluctant decision, as the gap between their capabilities and those of OpenAI's GPT was significant, necessitating a catch-up. Without open sourcing, the gap would only widen. Currently, the capabilities of open-source models and GPT are already quite small, and open sourcing now is more about seizing industry discourse power.
Jiang Daxin, founder and CEO of StepStar, stated that open sourcing aims to share the latest technological achievements. More importantly, multimodal models are a necessary path to achieving AGI, which is still in its early stages and requires developers to brainstorm together, expand the boundaries of model technology, and promote industrial implementation.
&#34;If you don't open source, you're out.&#34; An executive from a company providing AI cloud infrastructure told Wall Street Insight that around the Spring Festival this year, DeepSeek successively open-sourced foundational models and inference models, forcing other model companies into a corner where they couldn't survive without open sourcing or securing funding.
Several industry insiders indicated that the current strong advocacy for open-source models in the industry essentially stems from the fact that open source has become a core strategy for driving technological iteration, ecosystem building, and market expansion, while closed-source or &#34;pseudo-open-source&#34; models may gradually lose competitiveness due to technological barriers or ecological isolation.
However, open sourcing also brings commercial challenges. For many AI model vendors, the previous primary charging method was to price APIs based on tokens. Now that model weights are open-sourced, how to achieve sustainable growth in profitability has become an urgent problem to solve
An insider from an AI model vendor stated that this may force model companies to rethink their business models, shifting the company's commercial focus towards the service end. There are many issues to consider when deploying models; if applied in production, one must consider the stability and performance of product equipment, and the model requires many supporting toolchains. &#34;Providing enterprise-level services is where the commercial value lies in the future.&#34;
Liu Daoru believes that after open sourcing, the demand for model fine-tuning and distillation will also surge. Fine-tuning and distillation still have thresholds, and other model companies can assist enterprises in deploying in vertical scenarios. Additionally, different types of large models, such as multimodal large models and speech large models, are relatively suitable for other large model vendors due to their limited universality and high computing power requirements.
Currently, AI large model vendors have begun to focus on vertical applications in hopes of creating differentiated competitiveness. On March 3, AI unicorn Baichuan Intelligent initiated a round of layoffs, cutting its ToB business team in the financial sector, with employees signing departure agreements on the same day.
Wall Street Insights learned from within Baichuan that the optimization of financial business is aimed at concentrating resources, focusing on core advantageous businesses, and accelerating the realization of the vision of &#34;creating doctors, changing paths, and promoting medicine.&#34; It is understood that the AI pediatrician co-developed by Baichuan and Beijing Children's Hospital has officially &#34;started work&#34; in consultations with top experts on difficult cases.
AI Agents are considered the most important product form for AI deployment by 2025. On March 6, the world's first general-purpose AI Agent, Manus, quickly became a sensation across social circles, demonstrating excellent capabilities in writing articles, creating PPTs, and writing analytical reports. Although there are still many doubts, the popularity of Manus itself indicates that the market has high expectations for AI Agents.
Currently, companies like Zhipu, MiniMax, and Jietiao Xingchen are vigorously developing AI Agents. The enhancement of long-context capabilities and multimodal processing capabilities is a necessary condition for the development of AI Agents, which is also seen as the most important opportunity for AI startups moving forward.
In December last year, Zhipu disclosed new progress on Agents, with the release of AutoGLM, which can autonomously execute over 50-step long operations and can perform tasks across apps. AutoGLM is considered an important attempt towards an AI intelligent operating system. Recently, Zhipu reached a cooperation with Samsung, and in the future, will bring the Agent experience to Samsung's latest Galaxy S25 series phones.
Additionally, the Qianwen QwQ-32B model has integrated capabilities related to intelligent agents, enabling it to engage in critical thinking while using tools and adjust its reasoning process based on environmental feedback. The Alibaba Tongyi team stated that they will continue to explore the integration of intelligent agents with reinforcement learning to achieve long-term reasoning and explore higher intelligence, ultimately aiming for AGI.
After two years of competition in AI technology, the AI market is entering a new round of major reshuffling. Alibaba, Tencent, and ByteDance are emerging with their respective large ecological advantages. For AI startups, the future opportunity still lies in finding differentiated capabilities and continuously building competitive barriers
This will be an exceptionally brutal elimination round, and the gunfire has already sounded.
Risk Warning and Disclaimer
The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investing based on this is at one's own risk

DPSK

BABA

09988

On March 6th, Alibaba released a brand new reasoning model Tongyi Qianwen QwQ-32B, which performs comparably to DeepSeek-R1 and reduces deployment costs. This model is suitable for complex scenarios, and users can experience it for free through the Tongyi APP. Alibaba's stock price rose 8.39% in the Hong Kong stock market, reaching a nearly three-year high, with an increase of over 70% since the beginning of the year. Industry insiders are taking a wait-and-see attitude towards market feedback, as Alibaba faces challenges in the global AI competition

- Alibaba has launched the QwQ-32B inference model, which matches the performance of DeepSeek-R1 while significantly reducing deployment costs to consumer-grade graphics cards.  
- This model is designed for applications requiring rapid responses and high data security, and it has been made available for free through the Tongyi app and open-source platforms.  
- Alibaba's investment of over 380 billion yuan in AI infrastructure aims to enhance its cloud computing capabilities and support the growing demand for AI applications, marking a significant shift in the competitive landscape.

Alibaba Qwen is catching up with DeepSeek