# Chinese Open Source: A Definitive History

Kevin Xu published this in his newsletter Interconnected on 6 March 2026, the week the Chinese Premier's Two Sessions address named open source as a commitment for the next five-year plan. Xu writes as an insider. He led international expansion strategy at GitHub, ran go-to-market and community outside China for PingCAP, and advised Kyligence, and he discloses all three where they come up. His argument is that Chinese open-weight AI did not come from nowhere and was not a state plan. Unlike solar, semiconductors or batteries, open source in China was never declared strategic from the top. It grew for two decades as a subculture inside the tech industry, and the way DeepSeek, Qwen, Kimi and GLM release models is what that subculture had learned by the time LLMs arrived. Nathan Lambert lists it in the [[open-source-ai-reading-list]] as the history "leading up to AI".

Xu divides the story into six acts. The clipped copy lost the act headings and the timeline image is unreadable, so the division below follows the narrative rather than his labels.

## Takers: Linux and de-IOE

Linux reached China in the summer of 1994 with Gong Min, who happened to be doing research at the University of Helsinki when Torvalds released it. Shanghai had a Linux user group by 1997, Mozilla opened a China office in 2007, and the Linux Foundation held a summit in Guangzhou the same year. Xu dates the first real roots to the late 1990s, when a few large enterprises tried Linux to cut costs.

The founding story is Alibaba's "de-IOE" campaign. By 2008 Alibaba was Asia's largest database customer, running a 20-node Oracle RAC cluster that still could not keep up with Taobao. Jack Ma brought in Wang Jian from Microsoft Research Asia (a psychology PhD who, like Ma, could not code) with a budget of 1 billion RMB a year to replace IBM servers, Oracle databases and EMC storage with commodity x86 machines and MySQL. It took five years and a lot of resentment from engineers whose careers were built on IBM and Oracle certifications. The last IBM minicomputer went offline in May 2013, and Oracle left Taobao's core advertising system in July. The work seeded Alibaba Cloud and an internal database called OceanBase.

Xu stresses that nothing about this was nationalist. MySQL was by then owned by Oracle, the vendor Alibaba was escaping, and nobody treated that as a problem. JD.com adopted Kubernetes in early 2016, less than a year after Google released it, because it needed more scale than Walmart. The mindset of the era was "what open source can do for me".

## Community builders

Consuming open source is easy; collaborating in it takes norms, and Xu credits two institutions with teaching them. GitHub was the first. China was its second-largest country by users and activity by 2018, and it was never permanently blocked: the domain went dark briefly in 2013, the "Great Cannon" DDoS hit it in 2015, and each time developers protested because nobody could work without it. Xu believes it is still the largest Western site not technically blocked in China.

The second was Kaiyuanshe ("Open Source Society"), founded in October 2014 as a volunteer, vendor-neutral organization of individual members. It runs the annual COSCON conference, publishes a China Open Source Annual Report from GitHub and Gitee data, became the first Chinese member of the OSI, and in 2019 helped draft the Mulan Permissive Software License, the first OSI-approved bilingual Chinese-English license. It also ran the workshops that taught technically strong developers how to write a pull request and behave in code review. Xu's claim is that the polish with which the AI labs license models and manage their communities is an outcome of that work.

The subculture had its own holiday and its own protest. Programmers' Day on 24 October (1024) was chosen by an online vote in 2010 with no government involvement, in contrast to Russia's, which Medvedev signed into law. In March 2019 the 996.ICU repository, named for the 9am-to-9pm six-day week that lands you in intensive care, collected over 200,000 stars in weeks. Tencent's QQ and Alibaba's UC Browser blocked its website, and Jack Ma defended 996 as a blessing. The campaign produced the Anti 996 License, an MIT derivative requiring compliance with labor law, and in August 2021 the Supreme People's Court and the labor ministry ruled 996 illegal. Xu notes the irony that Silicon Valley startups now boast about 996, and admits the schedule never really died in China. The [[zai-playbook]] interview has a Z.ai product director working 18-hour days and trading jokes about an American "002".

From this Xu draws what he calls "open source ethos" (开源情怀): a belief in the positive-sum nature of open source, combined with a wish to disprove the Western assumption that anything Chinese is stolen, by building world-class technology and giving it away.

## Givers: startups and the giants

Flush Chinese venture capital in the mid-2010s followed Silicon Valley's enthusiasm for open-source companies. Apache Kylin (2015), out of eBay's China R&D lab, became one of the first Chinese Apache top-level projects and then the company Kyligence. TiDB, a Spanner-inspired distributed MySQL-compatible database from PingCAP, raised hundreds of millions and by 2018 was considered good enough to replace Oracle in China. Their success pushed the giants. OceanBase set a TPC-C world record in 2019 but stayed inside Ant Financial while TiDB grew in the open, so it was spun out in 2020 and open-sourced in 2021. Baidu released the Apollo self-driving platform in 2017, which still runs its robotaxis after 20 million rides. Alibaba's T-Head launched the Xuantie RISC-V processor project in 2021, and BYD opened its D++ vehicle platform in 2018. None of these made headlines, but Xu says they built the habits of community management and of balancing permissive licenses against commercial goals.

## Huawei and survival

Huawei arrived late. It was a top Linux kernel contributor by 2018 and content to work in other people's projects until the US Commerce Department blacklisted it on 15 May 2019, cutting it off from Google's Android services, American chips, and eventually TSMC. Huawei had been running "Plan B" projects since 2012: HarmonyOS from 2015, the HongMeng microkernel from 2016. It chose to open-source nearly all of it. openEuler came out on the last day of 2019, openGauss, openLooKeng and the MindSpore AI framework in 2020, and OpenHarmony in September 2020, used to seed the OpenAtom Foundation, China's first homegrown open source foundation with government backing. It kept contributing to Linux while doing this, and was the largest contributor to Linux 6.1 in December 2022. HarmonyOS NEXT in October 2024 contained no Android code and ran on HongMeng instead of Linux.

## The state arrives late

Made in China 2025, published in May 2015, does not mention open source once. That changed in 2021, when the MIIT's plan for the 14th five-year period used the term 27 times, held up OpenAtom as a model, and set a target of two or three open source communities with "global influence" by 2025. Xu calls the embrace double-edged. Being on the government's side is always better in China, and some local money did good. But in August 2020 the ministry had already anointed Gitee as the national code-hosting champion, so every notable project had to maintain a mirror there while the real activity stayed on GitHub and later Hugging Face. His summary: in open source, the bureaucrats followed the code and the community.

## The AI generation

What struck Xu about DeepSeek R1 in January 2025 was the launch rather than the benchmark: an MIT license, visible chain of thought "for all to see (and distill from)", and a follow-up open source week releasing tooling day by day. Qwen had released weights earlier under Apache 2.0. Moonshot founder Yang Zhilin and his team ran a Reddit AMA in English for Kimi's thinking model and later launched Kimi 2.5 in a direct-to-camera video on X; Z.ai does AMAs for each MIT-licensed GLM. When OpenClaw took off in early 2026, Kimi and MiniMax were among the first to offer it hosted, followed by Alibaba, Tencent and Baidu, which is the AWS and MongoDB playbook applied to AI. Xu leaves benchmarks and distillation to Lambert, pointing readers to pieces such as [[open-models-in-perpetual-catch-up]]; the distillation argument is on [[how-much-does-distillation-matter-for-chinese-llms]] and [[the-distillation-panic]].

He rejects the reading of all this as a state plan to flood the world with model weights as historically wrong. His alternative explanation is commercial. China's 2026 growth target of 4.5 to 5% is the lowest since 1991, going abroad with a "China" label is hard, and open source lets a product be judged on merit for free. For Chinese AI companies he calls it a necessity of overseas growth. He ties it to Jensen Huang's two favorite China talking points, that China has half the world's AI talent and that its open models are impressive, and argues they are one point: no country trains engineers without open source, and a country that does not promote it cannot hope to exercise sovereignty over AI.

## What the history does not settle

The link between the older community and the current labs is suggested more than shown. Xu asks whether Liang Wenfeng ever went to a 1024 party or Yang Zhilin to a Kaiyuanshe webinar, and leaves the questions open. He also sets DeepSeek aside as an anomaly with a hedge-fund origin, which removes the most prominent lab from the lineage he is drawing.

Later and earlier sources give narrower reasons for openness. His own [[chinas-structural-advantage-in-open-source-ai]] (June 2025) credits the tight link between Chinese academia and industry. In the [[zai-playbook]] (November 2025), Z.ai's Zixuan Li says the company kept its flagship closed in 2024 and opened it after R1 showed that American companies will not call a Chinese API but will download the weights. Lambert's [[notes-from-inside-chinas-ai-labs]] (May 2026) finds the labs practical rather than absolutist about openness. These fit Xu's commercial argument better than the ethos one. On the state, Lambert agrees that government help is real but its size unclear, and saw no sign of it steering technical decisions. Xu and Lambert later co-wrote [[banning-open-source-ai-would-be-a-mistake]].