聲明:本文來自于（ID:ykqsd.com）授權(quán)轉(zhuǎn)載發(fā)布。

農(nóng)歷春節(jié)前夕，深度求索公司旗下應(yīng)用程序和R1推理模型相繼登上App Store免費下載榜榜首，憑借優(yōu)質(zhì)的產(chǎn)品和合理的價格，引發(fā)了全民關(guān)注。

農(nóng)歷春節(jié)過后，阿里巴巴旗下的通義系列大模型不僅多次蟬聯(lián)“屠榜”佳績，還成為全球最大的開源模型，被開發(fā)者親切地稱為“源神”。

最新消息顯示，3月6日，阿里云發(fā)布了全新推理模型通義千問QwQ-32B，并正式開源。該模型采用更小的參數(shù)，在數(shù)學(xué)、代碼和通用能力等方面與深度求索的R1平分秋色。開源當(dāng)天，該模型即登上全球主流AI開源社區(qū)Hugging Face的趨勢榜。僅僅數(shù)日后，阿里云旗下的視覺基座模型萬相2.1也在Hugging Face趨勢榜和模型空間榜上登頂，成為近期全球開源社區(qū)最受歡迎的模型。

（千問QwQ-32B開源當(dāng)日即登頂Hugging Face趨勢榜，第四位為阿里旗下萬相2.1視覺模型。圖源|Hugging Face社區(qū)截圖）

這意味著，在全球前三的AI開源公司中，中國（杭州）已占據(jù)兩席。

從政策層面來看，人工智能正逐步進入國家發(fā)展議程。自2017年首次寫入政府工作報告，到2025年全國兩會，人工智能已連續(xù)7次被提及，且每次表述都更加具體化。一系列變化凸顯了中國對新一代人工智能發(fā)展的高度重視，同時也反映出一批中國企業(yè)在人工智能領(lǐng)域持續(xù)突破所形成的強大創(chuàng)新支撐力。

深度求索憑借開源性、性價比和降低算力依賴等優(yōu)勢，已成為近期最耀眼的明星企業(yè)。然而，單靠一己之力難以成林，要堅定中國人工智能發(fā)展的長期信心，就必須有更多持續(xù)性、全場景的突破。自2023年8月起，阿里巴巴持續(xù)堅持開源路線，累計開源超200款全尺寸、全模態(tài)模型的通義系列模型，似乎為近期的突破提供了最好答案。

《南華早報》今年2月報道稱，阿里巴巴通義千問系列模型使斯坦福、伯克利等高校能夠以低成本復(fù)刻深度求索的AI模型。文章指出，“阿里巴巴模型的能力再次證明，中國正在縮小與美國領(lǐng)先企業(yè)的人工智能差距，而基于阿里千問開源開放的路線，研究人員越來越多地利用阿里巴巴的技術(shù)來降低AI訓(xùn)練成本?！?/p>

那么，為何深度求索和通義系列模型會成為交相輝映的開源“雙子星”？開源為何成為中國AI破局的必然選擇？

中國開源“雙雄”

In the realm of AI competitions, the Silicon Valley narrative has almost concluded with victory, leaving little room for unexpected developments.

The closed-source model stifles technological innovation and knowledge sharing.

Power law scaling, exponentially elevating the difficulty of catching up with global leaders in computational power.

Monopolistic advantages enable businesses to extract excessive profits from users.

In the face of chip dependency, Chinese AI enterprises have achieved a two-generation technological edge in graphics processing units (GPUs). Continuing down the path of "large models = large computational power" leaves them in a passive chase.

By contrast, Chinese AI enterprises have embraced an open-source, distributed computing, and user-centric approach. This strategy marks a significant turning point in global AI competition, as noted by former Chief Executive of Google, Eric Schmidt.

The most widely recognized achievement of DeepSeek, its "breakthrough" status, can be attributed to three key factors: low operational costs, superior performance, and open-source accessibility. Specifically, in the post-training phase, DeepSeek-R1 leveraged advanced reinforcement learning techniques to achieve performance comparable to GPT-4, while requiring only 1/180th of the computational resources.

Furthermore, DeepSeek has adopted a completely free model in its applications, enabling it to rapidly climb the charts of free applications in multiple regions. Without any advertising, DeepSeek achieved 10 million active users in just seven days. While the exact timeline for achieving one hundred million users is not disclosed,瑞銀 analyst Lloyd Vowsil在研報中估計這一過程可能需要兩個月。

Though DeepSeek has seen steady growth in user base over the past year, its premium pricing strategy has limited its appeal to many potential customers. Currently, the Pro version requires a monthly subscription fee of 200 USD.

While ChatGPT has successfully attracted a sizeable user base through its free tier, its high subscription fee has excluded many users, as noted by the firm.

提一記的是，DeepSeek-R1同步開源了模型權(quán)重。DeepSeek在其開源倉庫統(tǒng)一采用標(biāo)準(zhǔn)化、寬松的MIT License，實現(xiàn)完全開源，不限制商用，且無需申請，還允許用戶通過蒸餾技術(shù)借助R1訓(xùn)練其他模型。

在DeepSeek系列模型名聲大噪之后，通義系列模型則逐漸被公眾發(fā)現(xiàn)，成為杭州AI雙雄共同構(gòu)筑開源界的中國宇宙，他們直接粉碎了開源模型性能不如閉源模型的論調(diào)。

從時間路線上看，早在2023年7月，阿里云首席技術(shù)官周靖人在上海世界人工智能大會上堅定表達了對開源路線的選擇，而通義系列模型在次月（2023年8月）就身體力行地開源了通義千問模型Qwen-7B，這也開啟了國內(nèi)巨頭企業(yè)開源大模型產(chǎn)品的先河。后續(xù)騰訊控股、智譜華章、百川智能等企業(yè)也陸續(xù)開源了多款大模型產(chǎn)品。

從開源數(shù)量上看，阿里已經(jīng)開源了Qwen、Qwen1.5、Qwen2、Qwen2.5等4代模型系列，覆蓋從0.5B到110B等的"全尺寸"，總計開源了200多個模型。相比DeepSeek開源的1.5B、7B、14B、32B、70B以及670B多個類型的模型，通義系列模型除了在尺寸上更多元外，還包括語音、視覺、文本等全模態(tài)。

從便捷性上講，雖然DeepSeek-V3、DeepSeek-R3模型可以實現(xiàn)本地化部署，但671B的滿血版DeepSeek-R1，需要8卡的服務(wù)器才可以部署，光硬件成本就在數(shù)百萬級。但上述阿里最新開源的千問QwQ-32B在個人用消費級顯卡NVIDIA 4090，甚至蘋果M4 Mac電腦上都可以運行。并且整體性能與DeepSeek-R1不相上下，使QwQ-32B在開源當(dāng)日就被開發(fā)者推上了Hugging Face趨勢榜榜首。

寬松的開源許可和部署條件，意味著只要開發(fā)者或企業(yè)愿意，就可以本地部署QwQ-32B，不花一分錢地使用高性能AI。開發(fā)者或企業(yè)還可以根據(jù)需求，用"蒸餾技術(shù)"去蕪存菁地保留需要保留的內(nèi)容，形成專用模型，讓它從事任何你希望的工作，比如司法、教育、醫(yī)療和情感陪伴，這些"蒸餾后"的模型甚至可以對外商用。

由于通義系列模型"全尺寸、全模態(tài)、全場景"的堅定開源策略，它也被多位產(chǎn)學(xué)研界的大咖專家給予了高度評價，并選擇其作為基座模型進行優(yōu)化和精調(diào)。

比如，李飛飛團隊以千問Qwen2.5-32B-Instruct開源模型為底座，訓(xùn)練出新模型s1，取得了與Open AI的o1和DeepSeek的R1等尖端推理模型數(shù)學(xué)及編碼能力相當(dāng)?shù)男Ч?；DeepSeek官方曾透露，其將DeepSeek-R1的推理能力蒸餾為6個模型開源給社區(qū)，其中4個模型是基于Qwen-32B蒸餾的模型；伯克利Tiny Zero及上海交大LIMO也都在通義系列模型底座基礎(chǔ)上激活了更強大的推理性能。

通義系列模型積極貢獻開源社區(qū)的同時，開發(fā)者和企業(yè)也在通過智慧"反哺"通義系列模型的進化和優(yōu)化。目前在開源社區(qū)Hugging Face上，通義系列衍生模型數(shù)量已突破10萬，成為全球最大的開源模型，持續(xù)領(lǐng)先于其他開源模型。今年2月，Hugging Face開源大模型榜單的前10名，全部是基于通義系列模型二次開發(fā)的衍生模型。

（Hugging Face開源大模型榜單的前10名全部來自通義系列模型的衍生模型。圖|Hugging Face截圖）

通義系列模型為何能被廣泛傳播？這背后離不開AI領(lǐng)域中開源與閉源兩大陣營的激烈競爭。

開放源代碼（Open Source）是指允許用戶獲取源代碼，進行任意使用、修改和學(xué)習(xí)。值得注意的是，自O(shè)pen AI發(fā)布GPT-3后，其后續(xù)版本如GPT-3.5和o1均采用閉源策略，這背后的原因包括安全、可控性、商業(yè)利益以及地緣政治等多方面的考量。

與之形成鮮明對比的是，Meta的"開源"采用的是更為嚴(yán)格的Meta Llama 3許可協(xié)議，這一協(xié)議下賦予用戶更多的限制權(quán)，具體內(nèi)容可通過下圖進一步了解。

（Meta、DeepSeek和通義系列模型的開源許可對比。圖|開源社區(qū)綜合整理）

其實，開源的力量不僅在于匯聚全球頂尖人才共同推進技術(shù)研發(fā)，更在于將技術(shù)成果普惠全球社會。中國制定的AI標(biāo)準(zhǔn)也為全球各國加快AI技術(shù)的普及提供了重要保障。這也促使全球各國加快AI技術(shù)的普及。

在正在進行的全國兩會上，許多海外記者也表示，他們國家的技術(shù)人員正在利用中國開源大模型"蒸餾"本國的模型，這體現(xiàn)了中國技術(shù)的影響力。

中國開源模型的速度促使全球AI企業(yè)加速創(chuàng)新，在農(nóng)歷春節(jié)后，各大科技巨頭紛紛出手：Open AI推出了o3-mini，提供免費使用；馬斯克推出了"最智能的人工智能" Grok3；Anthropic推出了混合推理模型Claude 3.7 Sonnet。

Open AI首席執(zhí)行官薩姆·奧特曼表示：沒有開源，就是站在了歷史的錯誤一邊。

Meta首席人工智能科學(xué)家楊立昆表示，與其說是中國打敗了美國的人工智能，不如說是開源戰(zhàn)勝了閉源。

阿里的新增長曲線

深度求索的橫空出世， challenge U.S. tech giants like Open AI, while Alibaba's Comprehensive Integriation Series Model has long ranked at the top of the world's largest open-source model repository, facts that alluded to the Chinese tech sector's recent tech lock.

Foreign investors show increased confidence in China, with Alibaba's stock price rising from 77.35 Hong Kong dollars per share on January 13 to 145.90 Hong Kong dollars per share at the peak on March 7, reflecting a rise of over 88.6%.

(Since January 13, Alibaba's Hong Kong-listed stock has surged by over 80%. Wind screenshot.)

The revaluation of Alibaba's value stems from its long-term investment in "AI + Cloud Computing" strategy.

Fifteen years ago, Alibaba made a strategic decision to invest in cloud computing research, and since 2018, it has been exploring AI-based large language models. Now, Alibaba not only holds the global leader and regional leader in cloud computing but also develops cutting-edge language models. Since February 2022, Alibaba has continued to heavily invest in cloud computing and AI, with Wu Yongming announcing plans to spend over 38 billion yuan over the next three years to build cloud and AI hardware infrastructure, exceeding the total investment of the past decade.

The widespread application of AI technology has driven the rapid growth of demand for cloud services across various industries, including AI technology products and public cloud services like data, storage, and computation. According to Alibaba Group's 2025財年 third-quarter financial report, the quarterly cloud service revenue returned to double-digit growth, reaching 31.74 billion yuan, while AI-related product revenue has remained stable for six consecutive quarters at the three-digit level. Additionally, Alibaba Cloud is the only Chinese cloud service provider that consistently maintains stable profitability.

Data shows that 80% of China's technology companies, 65% of China's專精特新 "small giants" enterprises, and 60% of China's A-share listed companies utilize Alibaba Cloud's computing services. Over 50 Chinese companies, including China National Oil Company, China State Grid, China United Bank, China中華 insurance, Hangzhou Metro Group, Meituan, and Meituan Tairou, are engaged in extensive cooperation in the application of deep computing and AI. Apple has chosen Alibaba to cooperate in China's C端 AI application market, signaling a swift completion of Alibaba's AI application ecosystem across all channels.

Currently, Alibaba Cloud operates 86 data centers in 28 global regions, positioning it as the global leader in cloud computing and the regional leader in China's cloud market, while also serving 5 million clients worldwide. In 2022, Alibaba Cloud first introduced the concept of Model as a Service (MaaS) in response to the era of AI, reconstructing a comprehensive cloud platform from hardware, computing, storage, networking, data processing, model training, and inference.

Bloomberg previously reported that Alibaba Cloud is playing a pivotal role in China's AI development and industrial upgrading.

As Alibaba exits the e-commerce stage, it is increasingly embedding itself into the broader context of China's AI-driven development and industrial transformation.

隨著AI開源和普及的加速，AI的應(yīng)用范圍逐步擴展到各個行業(yè)，這對推理算力的需求也提出了更高的要求，這為阿里云帶來了顯著的發(fā)展機遇。

DeepSeek與阿里云的開源合作與技術(shù)突破，深刻改變了中國AI發(fā)展的進程。而AI技術(shù)的驅(qū)動下，科技革命、資本熱浪以及產(chǎn)業(yè)升級，將對中國經(jīng)濟發(fā)展格局產(chǎn)生深遠(yuǎn)影響。

謝謝有你，支持原創(chuàng)！

一起轉(zhuǎn)發(fā)，讓更多人看到。

?智谷趨勢專注于服務(wù)中產(chǎn)階級的覺醒，助更多人實現(xiàn)財富價值。從宏觀經(jīng)濟到商業(yè)邏輯，從企業(yè)興衰到產(chǎn)業(yè)變遷，這里有最真實的中國，有最深刻的洞察，有最不可察覺的趨勢密碼。