A Field Guide to China's Open-Weight AI Ecosystem

Take the sceptic's best objection first, because it's a fair one. Building on an open-weight Chinese model means depending on a lab you can't ring up, under licences and export rules that can shift while you're mid-sprint, with training data nobody is going to show you. That's real, and no download chart makes it disappear. What the chart does is change the shape of the decision. This isn't one viral model you can wait out. It's a steady supply of them, from a whole row of labs, and the cost of ignoring the category goes up every quarter.

China's open-weight AI ecosystem is a cluster of labs, DeepSeek, Alibaba's Qwen, Moonshot's Kimi, Z.ai's GLM, ByteDance, MiniMax, Baidu and Tencent, that release powerful models with downloadable weights under several different licences. The release wave since June has moved the category on again. DeepSeek V4, Kimi K3, GLM-5.2 and MiniMax M3 all have 1-million-token context windows, and the past month alone brought Tencent's first flagship-scale open model and, on 8 August, open Qwen3.8 weights. Chinese open-weight models first overtook US ones in their share of Hugging Face downloads in 2025, 17.1% to 15.8% over the year to August in an MIT and Hugging Face study, and the gap has kept widening since. The more important August update is that several of these releases now target the same long-horizon coding and agent work as the leading closed models.

Key Takeaways- China's open-weight scene is a group of serious labs, not one viral model. The current flagships include DeepSeek V4, Qwen3.8, Kimi K3, GLM-5.2 and MiniMax M3, with Tencent's Hunyuan Hy3 new this summer.- It's strategy, not charity. Beijing treats open weights as an industrial asset, a thesis the US-China Commission calls the "two loops".- Adoption is already decisive: Chinese open-weight models overtook US ones in share of Hugging Face downloads in 2025, and by late February 2026 they were taking roughly 61% of token consumption inside OpenRouter's top 10.- "Open" no longer means one licence posture. Qwen3.6 and Tencent's Hy3 use Apache-2.0, DeepSeek V4 and GLM-5.2 use MIT, while the new Qwen3.8, Kimi K3 and MiniMax M3 use custom licences.- Hosted and downloadable releases still need separating. Alibaba's Qwen3.8-Max is a hosted flagship, but Qwen3.8's weights are downloadable under their own terms.

This is the field guide I wish someone had handed me before I started pulling these weights down and breaking things with them. The geopolitics sits one level up, in the pillar on whether China is winning the AI race. This piece stays at ground level.

Why did China open-source its AI?

China open-sourced its AI because open weights are a strategic asset, not a giveaway. The US-China Economic and Security Review Commission frames it as "two loops": a digital loop, where open weights seed adoption and iteration worldwide, and a physical loop, where cheap deployment across Chinese factories feeds proprietary industrial data back into better models. Policy first. Generosity a distant second.

The logic is straightforward once you see it. China accounts for 17.8% of global AI publications and 74.2% of granted AI patents in the 2026 AI Index. Open weights turn that research volume into reach. Give a capable model away and developers everywhere build on it, which spreads Chinese tooling, standards and dependencies far faster than any export deal could.

The political backing is explicit. Premier Li Qiang pledged at the 2025 World AI Conference that China would be "more open in sharing open-source technology and products", Xinhua reported, and the wider "AI+" initiative aims to fold AI into manufacturing, services and government, as the Carnegie Endowment's analysis of Chinese AI policy lays out. Open weights also route around chip export controls. If your edge is diffusion rather than owning the single best closed model, you don't need the newest accelerator to matter globally.

Here's the bit Western coverage tends to miss. Open-sourcing isn't the move of a country that's behind and hoping generosity buys it goodwill. It's closer to the opposite. When you can't win the closed-frontier race on compute alone, you change the game to one you can win: ubiquity. A free, good-enough, permissively licensed model that anyone can download and run on their own metal is a far stickier form of influence than a paid API nobody outside your borders can audit.

China open-sources its AI as deliberate industrial policy, not altruism. The US-China Commission's "two loops" thesis argues open models reinforce China's broader manufacturing and industrial dominance, while Premier Li Qiang's pledge to share "open-source technology and products", carried by Xinhua, and the "AI+" initiative back it politically.

The spend-and-strategy comparison against the US gets its own treatment in is China winning the AI race.

The Chinese open source AI models list: who builds them?

The Chinese open-weight scene runs on roughly nine serious players, and it's far more varied than the DeepSeek headlines suggest. Stanford's DigiChina brief on the ecosystem profiles four leading families in depth, Qwen, DeepSeek, Kimi and GLM, and name-checks a bench beyond them. Here's the who's who, with each lab's flagship open family.

Qwen still towers over the field on ecosystem size. By 21 January 2026 it had passed 1 billion cumulative downloads and more than 200,000 derivative models on Hugging Face, by Alibaba's own count; I'd treat that as a reported rather than audited figure, but an independent estimate of around 942 million Hugging Face downloads by March 2026 is close enough to confirm the scale. MIT Technology Review notes that Qwen variants alone account for more than 40% of new derivative models on the platform. Ecosystem size is no longer a proxy for the newest model, though. Alibaba released Qwen3.8's open weights on 8 August 2026, a 2.4T mixture-of-experts model with 95B active parameters, under a custom Qwen licence rather than Apache-2.0, while Qwen3.6 remains the current Apache-2.0 line.

Behind the main five, the bench matters more than it used to, and it's worth knowing by name rather than as a footnote. Tencent's Hunyuan Hy3 is the real news: released in July 2026 under Apache-2.0, it's the first time Tencent has put flagship-scale weights out in the open, and it lands straight into the serious evaluation pile. Baidu's open Ernie 4.5 family marked a genuine shift for a company that ran closed for years, though the open variants still trail the leaders on adoption and Ernie 5.0 has stayed closed. ByteDance is the odd one out: enormous consumer scale through Doubao and credible research checkpoints in Seed-OSS, but its flagship models remain closed, so check what's actually open before planning around it.

Two of these labs build their models a short ride from my apartment, and I still got this wrong at the start. I assumed it was a two-horse race between DeepSeek and Qwen, partly because that's the race the headlines cover, partly because in Hangzhou those are the two names you hear in every conversation. It isn't a two-horse race. In my own stack I keep reaching for GLM and Kimi on particular jobs, long-context work especially, where a smaller lab has quietly shipped the better tool. Read the ecosystem as "DeepSeek and the rest" and you'll miss half the good options.

Reference map: major Chinese open-weight labs and flagship model families. Compiled from Stanford's DigiChina brief and the labs' own release notes and model cards. Indicative, not exhaustive.

For a direct head-to-head between the leaders, see Qwen vs DeepSeek vs Llama.

China's open-weight ecosystem spans at least nine serious labs. Its August 2026 flagships include DeepSeek V4, Qwen3.8, Kimi K3, GLM-5.2 and MiniMax M3, with Tencent's Hunyuan Hy3 newly open under Apache-2.0. Stanford's DigiChina brief documents the field as diverse rather than a one-model story; the current model facts come from the labs' release notes and model cards.

Where can you find and run Chinese open models?

You find Chinese open models in the same places you find any open model. Hugging Face dominates as the host, with ModelScope as the Chinese-native hub. Adoption there isn't marginal. Chinese open-weight models overtook US ones in share of global Hugging Face downloads for the first time in 2025, 17.1% to 15.8% over the year to August in the MIT and Hugging Face study, and they haven't looked back. On the API side, Chinese models took roughly 61% of token consumption inside OpenRouter's top 10 in late February 2026 and were the majority of the whole platform's traffic by the summer, on OpenRouter's rankings. These weights sit on the mainstream infrastructure you already use.

There are three practical routes in. Hugging Face is where you download weights to self-host or fine-tune. ModelScope, Alibaba's hub, often carries Chinese releases early. Hosted APIs come either from the lab itself or from aggregators such as OpenRouter. Check availability per checkpoint: every flagship named above now has downloadable weights, while hosted flagships like Qwen3.8-Max sit on the labs' own platforms alongside them.

The licence picture matters more than people expect. Qwen3.6 and Hunyuan Hy3 use Apache-2.0. DeepSeek V4 and GLM-5.2 use MIT. Qwen3.8, Kimi K3 and MiniMax M3 carry their own licences. Those are meaningfully different starting points for a commercial review, and the terms can still change between checkpoints from one lab. Qwen is the live example: its open line moved from Apache-2.0 at 3.6 to a custom licence at 3.8 in a single release cycle.

My own workflow is boring on purpose. I pull the weights from Hugging Face or ModelScope, open the licence file before anything else, prototype against OpenRouter to see whether the model actually suits the job, then self-host the winner if the workload justifies the hardware. The licence check is the step people skip, and it's the one they come back to regret.

The current Chinese flagships are available through a mix of official APIs, Hugging Face and ModelScope. Licences differ sharply: Qwen3.6 and Hunyuan Hy3 are Apache-2.0, DeepSeek V4 and GLM-5.2 are MIT, and Qwen3.8, Kimi K3 and MiniMax M3 use custom terms. Check the exact checkpoint rather than relying on the family name.

If the governance questions are where you're stuck, my guide on whether DeepSeek is safe for enterprise walks through the diligence checklist.

What does this mean for Western teams?

For Western teams the headline has changed. Chinese open models are no longer only the cheap, good-enough option. Kimi K3, GLM-5.2, DeepSeek V4 and MiniMax M3 are designed for long-horizon coding, multimodal and agentic work, with 1M context across all four. The remaining questions are practical: which one fits the workload, whether its licence fits the product, and whether you want to carry the hardware bill for a checkpoint this large.

Think about what that changes in practice. For summarisation, extraction, classification, routing and drafting, you rarely need the single best frontier model in the world. You need one that's dependable, cheap, and yours to run where you like. Chinese open weights hit that target, and you can host them on your own machines instead of renting a closed API by the token.

The caveats are real and I'm not going to gloss them. Data governance, provenance, censorship behaviour and licensing all deserve proper scrutiny before anything touches production. That's a risk assessment rather than a veto. Carnegie's analysis is a useful reminder that these models arrive inside a policy context, so treating them as neutral commodities off a shelf is a mistake. Do the evaluation, then decide.

For Western teams, the practical meaning is that the world's default cheap, capable, permissively licensed model is increasingly Chinese and open-weight, with Chinese models leading Hugging Face download share and taking the majority of OpenRouter's top-10 token consumption by late February 2026. The right response is structured evaluation of governance and licensing, not reflexive avoidance.

On why these models are so cheap in the first place, I take the economics apart in why Chinese AI is cheap.

FAQ

What are the main Chinese open source AI models? The current leading open checkpoints are DeepSeek V4 Pro and Flash, Qwen3.8 and Qwen3.6, Kimi K3, GLM-5.2 and MiniMax M3, with Tencent's Hunyuan Hy3 new this summer. ByteDance's Seed checkpoints and Baidu's Ernie 4.5 broaden the field. Qwen remains the largest ecosystem by cumulative adoption, but its newest open weights carry a custom licence rather than Apache-2.0.

Why does China open-source its AI models? Because it's strategy, not charity. The US-China Commission's "two loops" thesis argues open models reinforce China's broader industrial dominance by spreading Chinese tooling and dependencies globally. It's backed politically by Premier Li Qiang's pledge to share "open-source technology and products", reported by Xinhua, and the "AI+" initiative described in Carnegie's policy analysis.

Are Chinese open models really open? The weights are downloadable for the releases named above, but the licences differ. Qwen3.6 and Hunyuan Hy3 use Apache-2.0, DeepSeek V4 and GLM-5.2 use MIT, and Qwen3.8, Kimi K3 and MiniMax M3 use custom licences. Open weights also do not mean open training data. Read the specific model card and licence before building.

Where can I download Chinese AI models? Hugging Face is the dominant host, with Alibaba's ModelScope as the Chinese-native hub that often gets releases first. If you'd rather not self-host, OpenRouter serves most of these models through one hosted API. Adoption on these platforms is now decisive: Chinese open-weight models overtook US ones in Hugging Face download share in 2025 and were the majority of OpenRouter's traffic by mid-2026, on OpenRouter's rankings.

The bottom line

China's open-weight ecosystem isn't one viral model. It's a deep bench of labs, DeepSeek, Qwen, Kimi, GLM, ByteDance, MiniMax, Ernie and Hunyuan, backed by deliberate national strategy. Beijing treats open weights as industrial policy, the "two loops" the USCC describes, and the adoption data already reflects it: Chinese open-weight models overtook US ones in Hugging Face download share in 2025, took roughly 61% of OpenRouter's top-10 token consumption by late February 2026, and were the platform majority by the summer.

The objection I started with still stands, and it should. You'd be depending on labs you can't call, under terms that can move, as Qwen's quiet licence change between 3.6 and 3.8 just showed. What the map buys you is the ability to be specific about which of those risks are actually yours to carry: for this workload, on this checkpoint, under this licence. That's a smaller and far more answerable question than "should we use Chinese models".

With the lay of the land in hand, compare the leaders head to head in Qwen vs DeepSeek vs Llama, work through the diligence in is DeepSeek safe for enterprise, or step back up to the pillar on whether China is winning the AI race. And if you'd like a second pair of eyes on a build, that's the kind of work I do.