The Open-Source Tide

In a wood-paneled hall on Chang’an Avenue in Beijing, on the morning of February 17, 2025, three weeks after the DeepSeek-R1 release had erased six hundred billion dollars of American market capitalization, Xi Jinping convened the first symposium on private enterprise the General Secretary had hosted in almost seven years. The seating chart was the message. In the front row, beside Jack Ma of Alibaba, Pony Ma of Tencent, Robin Li of Baidu, Lei Jun of Xiaomi, Wang Chuanfu of BYD, and Zeng Yuqun of CATL, sat a forty-year-old man in a dark zipped jacket and rimless glasses whose name almost no one in the West had heard before late January and almost no one inside the room had needed to know before that. Liang Wenfeng of DeepSeek had moved, in three weeks, from obscure quantitative-fund principal to seat assignment in the front row of a state symposium. The promotion was not subtle. The Chinese AI industry, in the General Secretary’s reading, had produced something that mattered, and the man who had produced it had been brought in front of the cameras to be acknowledged.

What the cameras did not show was the ecosystem behind him. The Western press had described DeepSeek as a single point in space, a Hangzhou hedge-fund anomaly that had emerged from nowhere, trained a frontier model on mistreated H800s, and disappeared back into the fog. The framing was useful for politics and bad for analysis. By February 2025, DeepSeek was the most visible node in a Chinese open-source AI ecosystem that had, over the previous eighteen months, become the largest concentration of openly licensed frontier and near-frontier models in the world. By the time Hugging Face published its annual state-of-open-source report in early 2026, Chinese organizations accounted for roughly forty percent of monthly model downloads on the platform. Alibaba alone had produced more derivative model weights than Google and Meta combined. The Llama family that Mark Zuckerberg had positioned in 2023 as the canonical open-weight foundation had, by mid-2025, fallen to roughly fifteen percent of new derivatives. Qwen had passed forty.

The lab that had built that family sat inside the largest e-commerce conglomerate in China. Alibaba’s Tongyi Qianwen, branded internationally as Qwen, had begun in 2023 as a research effort under Cloud Intelligence Group, the cloud arm Joe Tsai’s restructuring had pulled into the company’s strategic core. The Qwen team’s first major open release, Qwen-7B, had landed in August 2023, modest by the standards Meta had set with Llama 2 a month earlier and aggressive by the standards of Chinese open releases, almost none of which until then had carried fully permissive licenses. By September 2024 the team had released Qwen2.5, a family of dense and mixture-of-experts models trained on roughly eighteen trillion tokens, with sizes from half a billion to seventy-two billion parameters and specialist variants for coding and math. In November 2024, three weeks before DeepSeek-V3 reached arXiv, came QwQ-32B-Preview under Apache 2.0, an open reasoning model that sat within striking distance of OpenAI’s o1-preview on math and coding benchmarks and that, unlike o1, anyone could download and modify. On April 29, 2025, Alibaba shipped Qwen3, trained on thirty-six trillion tokens, with a 235-billion-parameter mixture-of-experts flagship that activated twenty-two billion per token and hybrid reasoning support that let a single model choose, on the user’s instruction, whether to think step by step or answer immediately. By the autumn of 2025, Qwen variants and their fine-tunes were the most-downloaded open models on Hugging Face, accounting for an order of magnitude more derivative repositories than the second-place family.

The pattern was not a one-company phenomenon. Across the autumn of 2024 and the year that followed, a sequence of Chinese labs that had been quiet through the early ChatGPT era began shipping open weights at a tempo no Western lab matched. In Beijing, Zhipu AI, the Tsinghua-affiliated company that had grown out of the GLM research line, released GLM-4 in early 2024 and then, on July 28, 2025, GLM-4.5, a 355-billion-parameter mixture-of-experts model with thirty-two billion active parameters, paired with a smaller GLM-4.5-Air variant at one hundred and six billion. Both were released under MIT-style terms. Zhipu had spent 2024 and 2025 quietly recruiting senior engineers out of the larger Beijing technology firms, and by the time of the GLM-4.5 release was widely understood to be the lab the Chinese government would back if it wanted a state-aligned analogue to OpenAI. The company began trading externally under the domain z.ai later that year, an English-language facade meant to make the international rollout less audible to American export-control watchers.

In Hangzhou, on the same Huanglong Road corridor that housed DeepSeek, a four-year-old startup called Moonshot AI had been operating in adjacent water. Yang Zhilin, a Tsinghua and Carnegie Mellon graduate who had cofounded the company in 2023 with two former classmates, had spent the early scaling years emphasizing context length. Moonshot’s product, Kimi, had begun in early 2024 with a 200,000-character window that the company expanded, in March, to two million characters in a closed beta that turned its consumer chatbot into the most-discussed Chinese AI app of the spring. By July 2025 the company had open-sourced Kimi K2, a one-trillion-parameter mixture-of-experts model with thirty-two billion active parameters, posting the highest open-source coding scores any Chinese model had reached. The bet on long context was well suited to the agentic work the Chinese product market was beginning to demand: models that could read entire code repositories, full legal contracts, or week-long conversation histories in a single prompt and reason coherently across them.

In Shanghai, MiniMax, founded in 2021 by former SenseTime researcher Yan Junjie, took a third route. The company had spent its early years building proprietary chat products, with its flagship Talkie aimed at consumers and its enterprise tier at game studios and content companies. In January 2025, two weeks before R1, MiniMax open-sourced the MiniMax-01 family, a 456-billion-parameter mixture-of-experts model whose distinguishing feature was a hybrid attention scheme combining lightning attention and a small fraction of standard softmax attention. The design let the model handle four-million-token contexts without the memory blow-up pure softmax attention produced at long sequence lengths. MiniMax-M1, mid-2025, was a hybrid-attention reasoning model the company claimed needed roughly seventy percent less compute than peers on long-context reasoning. MiniMax listed on the Hong Kong Stock Exchange in January 2026 and shares doubled on the opening day. The IPO confirmed what the open-source releases had implied. Beijing, which had spent 2021 through 2023 cracking down on the consumer-internet sector, had quietly reversed posture for AI. Companies that produced models the world wanted to download could go public, and the IPO desk could clear the listing.

Beneath the four marquee labs sat a longer tail. ByteDance had spent 2024 building Doubao into China’s most-used consumer chatbot, with its February 2026 Lunar New Year peak crossing one hundred million daily active users, ahead of every Western competitor. Doubao itself remained closed, but ByteDance’s research arm shipped open-weight Seed models in 2025 that were respected by practitioners and ignored by the press. Tencent’s Hunyuan family had stayed largely closed through 2024 before releasing dense and mixture-of-experts variants in 2025 under restrictive licenses. Baidu, which had bet most heavily on its own foundation models, finally yielded to the open-source momentum and released ERNIE 4.5 under a permissive license in June 2025, a ten-variant family running from a 0.3-billion-parameter dense model to a 424-billion-parameter mixture-of-experts. The release, after years of Robin Li arguing in keynotes that open-source AI was strategically self-defeating, was read inside the industry as the moment Baidu had given up its earlier framing rather than continue to lose ground. By the November 2025 Baidu World conference, Li was on stage rolling out ERNIE 5.0, a 2.4-trillion-parameter native multimodal model whose pricing undercut OpenAI’s by an order of magnitude.

Around the major players, a denser layer of specialists filled in. Step, based in Shanghai under the former Microsoft Research Asia executive Jiang Daxin, had built a reputation for open multimodal models. Baichuan, founded by former Sogou builder Wang Xiaochuan, had shipped competitive mid-sized open models through 2024 before pivoting in 2025 toward enterprise applications. Zero One, the company Kai-Fu Lee had founded in March 2023 and that the Western press had treated as the most natural China-side analogue to OpenAI, had spent 2024 trying to scale its Yi family to the frontier and had concluded, by the spring of 2025, that the cost was beyond what the company could fund. In March 2025 it stopped pre-training new foundation models and pivoted to enterprise services built on top of DeepSeek and Qwen, with Lee telling investors that competing with the well-resourced labs at the frontier was no longer the right business. Even with state and venture capital flowing freely, building a frontier model required a budget the smallest of the named labs could no longer assemble.

What the surviving labs did with their budgets was the part that mattered. The architectural innovations that had let V3 train an order of magnitude cheaper than its American peers, multi-head latent attention, fine-grained mixture-of-experts routing, and FP8 training stacks that had been considered too noisy for production scale, were public knowledge inside the ecosystem within weeks of the V3 paper. Qwen3’s mixture-of-experts variant adopted some of them. GLM-4.5’s architecture explicitly cited V3. Kimi K2’s training stack borrowed others. The labs shared engineers across job changes, papers across arXiv, and benchmarks across release cycles. Western researchers, accustomed to the closed-development culture of OpenAI and Anthropic, found themselves reading three or four Chinese papers a week describing techniques the American labs would not have published if they had developed them. The Chinese labs published because publishing was free advertising for the recruiters and because secrecy, in a country whose state had already announced AI as a national priority, was both impossible and pointless. A technique invented at DeepSeek in November showed up in Qwen by April and in MiniMax by July.

The geography mattered too. Hangzhou, where DeepSeek and Moonshot operated, had become the most concentrated cluster of Chinese AI talent outside Beijing, drawing on Zhejiang University’s electronic engineering program, Alibaba’s headquarters, and a generation of returnee engineers from Silicon Valley. Beijing held the policy weight, the state research institutes, and the campuses of Tsinghua and Peking University that fed Zhipu. Shanghai held MiniMax, Step, and the cluster around Fudan University. Shenzhen held Tencent’s research arm, the hardware ecosystem that ran from Huawei downward, and a thicker layer of robotics and embodied-AI labs that benefited from the Pearl River Delta’s manufacturing density. The four cities, with their constant exchange of researchers, formed a Chinese AI corridor whose total population of senior researchers approached, on credible Chinese government counts, roughly half the working population of the equivalent Bay Area cluster. The deep gap was no longer in talent. It was in the chips that talent could deploy.

The chip gap was the part the open-source pattern had developed against rather than around. With the exception of the largest hyperscalers, the Chinese labs did not have access to H100-class clusters at the scale OpenAI or Anthropic could marshal. They had, mostly, H800s that had cleared customs before the October 2023 thresholds bit, A800s and earlier-generation A100s acquired through routes the labs were careful not to describe in print, an emerging stockpile of Nvidia H20 inference accelerators sold legally under the post-2023 perimeter, and a slowly rising domestic supply of Huawei Ascend 910B and 910C parts. The total fleet across the ecosystem ran, on SemiAnalysis estimates, to several hundred thousand training-class accelerators. The American hyperscalers were on track to deploy several million by the same date. The factor of ten in raw compute had selected, more than any deliberate strategic choice, for what the Chinese labs were good at: architectural efficiency, low-level systems work, PTX-level kernels, training-stack optimizations whose returns scaled with the constraints they were applied against. The constraint had, paradoxically, made them better at the kind of engineering the American labs had stopped doing because Nvidia’s compilers had been good enough.

The release strategy compounded the engineering. By publishing weights under permissive licenses, the Chinese labs propagated their work to anyone with a credit card and a workstation. Russian researchers fine-tuned Qwen for Russian-language tasks. Brazilian and Argentine groups quantized DeepSeek to run on consumer hardware. Gulf labs trained sovereign-AI products on top of GLM and Qwen. African universities downloaded ERNIE for academic research. Each downstream user was an ambassador for the underlying model and an entry on the cost ledger Beijing had been quietly building for the export-control architecture. By April 2026, when V4 landed, Chinese open source had achieved what no Chinese commercial product had managed in any prior technology cycle: it had become a meaningful default for users in countries where American sanctions made closed American models difficult to acquire. The diffusion was real. It was also bounded. The frontier of capability remained, in early 2025, at OpenAI, Anthropic, and Google. The Chinese open-source labs were closing on it, but they were closing from behind.

What none of this meant was that Chinese AI had become a coordinated state apparatus. The labs competed bitterly on benchmarks, talent, government contracts, and the limited supply of high-end domestic chips. DeepSeek and Moonshot, neighbors on Huanglong Road, had been quietly poaching engineers from each other across 2024 and 2025. Alibaba and Tencent had run parallel and largely uncoordinated cloud strategies. ByteDance had refused to license its Doubao technology to other Chinese firms even when state mediators had tried to arrange it. Beijing had not, by any credible reporting, picked a single winner. What it had done was provide capital, regulatory cover, and political signaling broad enough to keep the ecosystem competitive without naming a sole national champion. The model was less Soviet planning and more Korean chaebol-era industrial policy: pour money into a defined sector, accept that some firms would fail, and let the market sort the survivors. By February 2025, with Liang in the front row and Xi above the seating chart, the survivors had begun, visibly and with state encouragement, to sort.

The American policy community read the symposium two ways. The export-control hawks read it as evidence that the regime was working, that DeepSeek’s existence under sanction was a closing-of-the-gap story rather than a leapfrog story, and that the right response was to tighten further. The skeptics read the same evidence in the opposite direction. Chinese open-source had, on Hugging Face’s own data, become the global default for the long tail of fine-tuning, distillation, and on-device deployment. The American closed-weights labs faced a competitor whose products were free, whose languages were broader, and whose cost-per-token undercut OpenAI’s by an order of magnitude. The export controls had constrained the frontier of Chinese training. They had not constrained the diffusion of Chinese models around the world, and the diffusion was accelerating.

Liang said almost nothing publicly through the spring of 2025. His only on-record contribution to the debate was a sentence he had given a 36Kr interviewer the previous July: money had never been the problem, the bans on advanced chips were. By February the line had aged into something less plaintive. The bans had not stopped him. They had, on the evidence of the seating chart and the Hugging Face download numbers and the release cadences of Qwen and Kimi and GLM and MiniMax, produced a country whose open-source AI ecosystem was, in eighteen months, the largest in the world. What Beijing would do with that ecosystem, and what Washington would do in response, were questions the next turn of the rule cycle would have to answer.