導(dǎo)語：

近日，我在《金融時報》的專欄文章里，探討了中美大模型領(lǐng)域的新競爭格局。

當(dāng)美國科技巨頭們在封閉的實驗室里，以“贏者通吃”的寡頭競爭模式建立先發(fā)優(yōu)勢時，一股強(qiáng)勁的中國開源力量正迅速崛起。中國大模型采取的并非單打獨斗的路線，而是攜手共進(jìn)、集體攻堅的模式，走出了一種具有中國特色的“破局范式”。

這場開源與閉源的博弈，酷似當(dāng)年智能手機(jī)操作系統(tǒng)領(lǐng)域的安卓與iOS之爭。中國AI正憑借其“安卓策略”，以開放、可定制化等優(yōu)勢，撬動并重塑著全球AI的格局。

以下是專欄文章全文：

中國大模型不是依靠單打獨斗，而是攜手共進(jìn)以集體力量攻克課題。

作者：李開復(fù)

零一萬物 CEO、創(chuàng)新工場董事長

今年一月，中國大模型公司“深度求索”（DeepSeek）發(fā)布其 R1 大語言模型時，美國納斯達(dá)克指數(shù)曾出現(xiàn)單日 3% 的下跌。這款模型在性能上足以匹敵當(dāng)時市場領(lǐng)先的所有美國閉源大模型，而其所消耗的算力卻僅是后者的一小部分，這無疑暗示著美國在生成式 AI 領(lǐng)域的先發(fā)優(yōu)勢正在收窄。更引人注目的是，R1 模型選擇了開源，任何人都可以免費下載并用于商業(yè)用途。

時至今日，我們有更充分的理由相信中國大模型公司能夠與美國優(yōu)秀同行全球競逐。深度求索最新推出的兩款模型，其推理能力已經(jīng)媲美 OpenAI 的 GPT-5 和谷歌的 Gemini-3 Pro。繼 R1 和阿里巴巴的通義千問（Qwen）取得巨大成功后，開源模型在中國正逐漸成為行業(yè)主流。百度、智譜、月之暗面和美團(tuán)等公司均允許用戶下載、研究并定制化使用它們的頂尖開源模型。這種開源開放的做法與美國大模型領(lǐng)域日趨封閉的開發(fā)模式形成了鮮明對比，也為中國大模型行業(yè)提供了一種獨特的、有中國特色的破局范式。

開源模型賦予了用戶按需定制模型的能力。例如，可以針對垂直行業(yè)進(jìn)行微調(diào)。更重要的是，模型可以做企業(yè)私有化部署，無需將數(shù)據(jù)傳送給第三方服務(wù)器。對于科研人員、高校師生、技術(shù)愛好者和創(chuàng)業(yè)者而言，免費的開源模型也大幅降低了其獲取前沿AI技術(shù)的上手門檻。

我個人對此深有體會，因為早在 1988 年，在我完成語音識別領(lǐng)域的博士論文后，我在卡耐基梅隆大學(xué)（CMU）的導(dǎo)師、圖靈獎得主 Raj Reddy 教授就建議我將工具包進(jìn)行開源。幾十年過去了，這個工具包仍被開源社區(qū)使用和更新 [1]。這讓我深切體會到了開源社區(qū)的力量以及共享資源的持久生命力。

CMUSphinx開源項目：

https://cmusphinx.github.io/

隨著越來越多的中國大模型公司選擇開源自身模型，整體產(chǎn)業(yè)發(fā)展進(jìn)程正在顯著加速。不同公司的工程師得以相互研究彼此的模型，以及在此基礎(chǔ)上獨立開發(fā)和迭代出數(shù)千個變體。這使得創(chuàng)新者們能夠篩選整合不同模型的優(yōu)點，進(jìn)行漸進(jìn)式改良。這就好比拔尖的學(xué)生們不僅僅依賴于個人智力，反倒齊心組成了共學(xué)小組來一起攻克課題。這使得中國大模型企業(yè)有更多的杠桿去與美國同行競爭。

01

必然的選擇：

中國的“后發(fā)追趕”戰(zhàn)略

中國選擇開源道路實屬必然。盡管 Meta 在借助 Llama 模型推動模型開源，但大多數(shù)美國開發(fā)者仍將最尖端的大模型視為私有資產(chǎn)。究其根本，美國在生成式 AI 領(lǐng)域的早期領(lǐng)先地位，是以經(jīng)典的“硅谷模式”建立起來的：OpenAI、Anthropic 和 xAI 等公司依托風(fēng)投的巨額投資，采購高性能 GPU，并招募頂尖研究人才，在封閉的實驗室中開發(fā)模型。他們?nèi)缃裾萑胍粓觥摆A者通吃”的寡頭競爭，各家都寄希望于打造性能最強(qiáng)的模型，擠壓對手的空間，以此來建立市場壟斷地位。

處于追趕狀態(tài)的中國大模型行業(yè)，則必須將重點放在“多快好省”的效率上，著力開發(fā)算力需求更低、使用成本更優(yōu)的模型，以生態(tài)打法走變道超車。例如， DeepSeek 免費開源模型，正是為了鼓勵用戶基于 DeepSeek 基座模型構(gòu)建生態(tài)系統(tǒng)。在 R1 模型發(fā)布后數(shù)周內(nèi)， Hugging Face 上的開發(fā)者就創(chuàng)建了超過 500 個衍生模型，累計下載量達(dá) 250 萬次 [2]。

圖源：OpenLM.AI（2025年12月11日）

目前，全球排名前十的開源模型 [3] 幾乎全部來自中國（截至2025年12月11日）。這一優(yōu)勢已如此明顯，以至于谷歌前 CEO 埃里克·施密特 [4] 曾提出警示，開源模型領(lǐng)域，美國人工智能企業(yè)可能將完全被中國AI反超。

02

大模型領(lǐng)域的

“安卓 vs iOS”博弈

當(dāng)然，這一切并不意味著中國必然會在人工智能競賽中超越美國。美國公司在研究領(lǐng)域仍保持領(lǐng)先，并在持續(xù)投入海量資源進(jìn)行開發(fā)。他們的企業(yè)客戶愿意支付高昂訂閱費來使用閉源模型，為下一代的模型研發(fā)提供真金白銀。此外，美國公司可以暢通無阻地獲取世界一流的 GPU，這仍是生成式 AI 迭代不可或缺的硬件基石。

因此，未來人工智能大模型的發(fā)展格局很可能類似于智能手機(jī)操作系統(tǒng)領(lǐng)域的 Apple 與 Google 之爭：美國公司正在構(gòu)建一個類似 Apple iOS 的封閉生態(tài)系統(tǒng)，提供高價的“精品服務(wù)”；而中國公司的發(fā)展路徑則更接近 Google 安卓（Android），打造開放且可定制化的操作系統(tǒng)。

圖源：pexels

雖然 iPhone 在高消費力群體中廣受歡迎且利潤豐厚，但覆蓋更廣的安卓系統(tǒng)卻成為全球 70% 以上智能手機(jī)的操作系統(tǒng) [5]。中國的人工智能企業(yè)正遵循類似的“安卓策略”，通過技術(shù)開放與合作，推動產(chǎn)品與服務(wù)的豐富性和覆蓋率，以實現(xiàn)更大版圖的行業(yè)競爭力和影響力。

本文翻譯自《金融時報》英文專欄，原文如下：

China's open-source Al is a national advantage

The models are akin to studying together to ace a test instead of relying on individual knowledge

BYKAI-FU LEE

When Chinese artificial intelligence company DeepSeek released its R1 large language model in January, America's Nasdaq index fell 3 per cent in one day. The model rivalled market-leading US AI models in performance while using a fraction of their computing power, suggesting that America's head start in generative AI might be shrinking. What's more, it was made available open-source. Anyone could download it free and adapt it for their own commercial use.

Today, there is more reason than ever to believe Chinese AI companies can rival their US peers. DeepSeek's latest two new models match the reasoning performance of OpenAI's GPT-5 and Google's Gemini-3 Pro. The runaway success of R1 and Alibaba's Qwen have made open-source models the norm in China. Companies like Baidu, Zhipu, Moonshot AI and Meituan all allow users to download their cutting-edge models, interrogate how they work and adapt them. Contrasted with the secretive development of LLMs in the US, they offer a distinct Chinese pathway for progress in AI.

Open-source AI gives users the ability to customise models — fine tuning them for use in a specific industry, for example. The models can also be run on a customer's internal servers, which means corporate users don't have to send their data to AI companies. And free, open-source models make state-of-the-art AI affordable for researchers, students, hobbyists and entrepreneurs.

I know this from personal experience. Back in 1988, after finishing my PhD on speech recognition, my adviser, Turing Award recipient Professor Raj Reddy, suggested I open-source the toolkit. Decades later, it is still being used and updated [1]. This has shown me the power of open-source communities and the longevity of a shared resource.

As more Chinese AI companies have open-sourced their models, development has accelerated. Engineers at different companies study each other's models as well as the thousands of variants developed independently, allowing innovators to cherry pick features and make incremental improvements. The effect is akin to studying together to ace a test, rather than relying on individual intelligence. Today, there is more reason than ever to believe Chinese AI companies can rival their US peers.

This was born from necessity. While Meta promotes an open-source approach to AI through its Llama model, most US developers keep their cutting-edge LLMs to themselves. America's early lead in generative AI was developed in classic Silicon Valley fashion. Companies like OpenAI, Anthropic and xAI used huge quantities of venture capital to procure high-performing graphic processing units (GPUs) and the best researchers to develop models in closed labs. They are now engaged in a winner-takes-all battle to build the best-performing model, squash competition and establish a monopoly.

Forced to play catch-up, China's AI industry has focused on efficiency, developing models that require less computing power and so are cheaper to use. DeepSeek chose to give away its model to encourage customers to build an ecosystem of products on top of it. Within days of the release of its R1 model, for example, developers on AI community Hugging Face had created more than 500 derivative models, which were downloaded 2.5mn times [2].

Today the 10 top-ranked open-source AI models [3] are almost all Chinese. The dominance is now so pronounced that former Google CEO Eric Schmidt [4] has warned that US companies risk ceding open-source AI to China completely.

None of this means China will necessarily win the AI race against the US. American companies continue to lead in research and to pour huge resources into development. Their corporate customers are willing to pay high subscription fees to access closed models, thus funding further R&D. Unlike Chinese businesses, which face US export restrictions on Nvidia chips, they also enjoy unencumbered access to the best-in-class GPUs — a hardware essential to AI computation.

The future of AI development could therefore resemble the rivalry between Apple and Google in smartphone operating systems. Like Apple's iOS, US companies are building a closed ecosystem, charging high prices to access a premium product. China's AI approach is closer to Google's open and customisable Android operating system.

While iPhones are popular with wealthy consumers and highly profitable, Android powers over 70 per cent of smartphones globally [5]. China's AI companies are following a similar "Android strategy", aiming for broader reach through open technologies.

[1] https://cmusphinx.github.io/

[2] https://x.com/ClementDelangue/status

/1883946119723708764?ref_src=twsrc%5Etfw

[3] https://openlm.ai/chatbot-arena/

[4] https://www.ft.com/content/84cf0b2e-651d-4cb4-b426-ebc7afd634fa

[5] https://gs.statcounter.com/os-market- share/mobile/worldwide