Is China winning open-source AI?
In our updated market reports, you will find everything you need
SUMMARY
Yes, China is currently winning open-source AI at the model layer, although “open-weight AI” is the more accurate description of what it leads.
The lead is no longer a DeepSeek story. GLM, MiniMax, Qwen, MiMo, Kimi and Tencent have created enough depth that a weak release from one lab no longer changes the national picture.
China’s strongest evidence is the combination of quality and breadth. Seven of the nine open-weight systems scoring 40 or above on the current Artificial Analysis table come from Chinese labs.
Usage is beginning to support the benchmark results. Chinese models have temporarily overtaken American models in observed OpenRouter token volume, and several now handle large coding, agentic and general-purpose workloads rather than collecting downloads alone.
Price is the force multiplier. Chinese models do not need to beat Claude or GPT on every difficult task when they can complete routine work at a fraction of the cost and hand only the hardest cases to a premium model.
The word “open” still needs discipline. Most leading releases provide downloadable weights, code and deployment instructions, but not the full training data and process required for fully reproducible open-source AI.
Self-hosting is also less democratic than it sounds. Frontier Chinese models can require hundreds of gigabytes or more than a terabyte of weights, so most companies will use managed endpoints, quantized versions or smaller models.
China leads the weights, but it does not control the whole ecosystem around them. American clouds, Nvidia’s software stack, Hugging Face, vLLM and other foreign platforms still shape discovery, deployment, billing and customer relationships.
Chip restrictions are raising China’s costs without stopping its progress. They have pushed Chinese labs toward sparse architectures, lower active-parameter counts and aggressive efficiency work, while domestic hardware still carries a noticeable software and reliability tax.
The likely outcome is a split market rather than a clean national victory. China is ahead in downloadable models and cost-efficient workloads; the United States remains stronger in closed frontier systems, cloud distribution, accelerator software, capital and advanced compute.
China’s lead is real, but it is not permanent. One major American release could retake the top benchmark position quickly; matching China’s current range of credible model families would take longer.
Is China winning open-source AI?
Why does China look so strong in open-source AI right now?
China looks strong in open-source AI today because several Chinese labs are releasing competitive models at the same time, and developers are actually using them.
DeepSeek changed how people viewed China, but the recent shift is much broader. Hugging Face found that most newly created models that trended during 2025 were either built in China or derived from a Chinese model. Baidu went from no releases on the platform in 2024 to more than 100 the following year, while ByteDance and Tencent increased their releases roughly eight to nine times. Companies that had preferred closed models, including Baidu and MiniMax, also began publishing weights.
Usage has moved even faster. OpenRouter reviewed more than 450 trillion tokens generated this year and found that Chinese models overtook American models in token volume for a period earlier in the year. DeepSeek V4 drove much of that jump, but MiniMax, Xiaomi, Alibaba, Moonshot AI and Z.ai were also gaining traffic.
OpenRouter attracts developers who compare models, switch providers and care about price, so its numbers are not global market share. Still, the direction is hard to dismiss. China now has repeated releases, several credible laboratories and real workloads moving onto its models. That combination did not exist when DeepSeek first became famous.
What does “winning open-source AI” actually mean?
Winning open-source AI currently means leading the downloadable model layer on quality, adoption, cost and ecosystem influence.
A country can lead this particular contest without controlling every chip, cloud platform or consumer chatbot. The lead comes from models that developers can download, adapt and deploy without accepting a large performance penalty, backed by fine-tunes, quantizations, hosting support and regular usage.
Three larger claims need to stay separate. Open-weight leadership says little by itself about who leads artificial intelligence overall, whether the models meet the strictest definition of open source, or which companies capture the revenue around them.
We can judge the race with four questions. Who releases the strongest weights? Which model families keep developers after the launch excitement fades? Who gives users a clear economic reason to switch? And whose architectures are becoming common building blocks?
On those questions, China is ahead. The United States remains stronger in frontier closed models, cloud distribution and advanced chip infrastructure.
If you want more recent data on this point, please see our latest foundation model market report.
Are Chinese open-weight AI models the best in the world today?
Chinese labs currently produce most of the strongest open-weight AI models, with Meta emerging as the only serious challenger at the very top.
Artificial Analysis currently gives GLM-5.2 and Meta’s Muse Spark 1.1 the highest open-weight Intelligence Index score, at 51. MiniMax-M3 and DeepSeek V4 Pro follow at 44, while an earlier Muse Spark version scores 43. Xiaomi’s MiMo-V2.5-Pro and Moonshot AI’s Kimi K2.7 Code both reach 42, Tencent’s Hy3 reaches 41 and DeepSeek V4 Flash reaches 40.
We counted nine open-weight systems scoring 40 or above on the live table. Seven come from Chinese labs. Meta supplies the other two. That is depth across Z.ai, MiniMax, DeepSeek, Xiaomi, Moonshot AI and Tencent, not one company briefly reaching first place.
A recent assessment by the U.S. Center for AI Standards and Innovation points in the same direction. Its evaluators called GLM-5.2 probably the most capable open-weight model available when it was released and found its broad capability similar to GPT-5.2, a closed American model released several months earlier.
Benchmarks can be noisy, and the Stanford AI Index warns that some popular tests contain high error rates or reward models that have adapted to the benchmark. No single score settles the race, but Chinese models keep appearing near the top across independent evaluations.
| Open-weight model | Lab | Country | Artificial Analysis score |
|---|---|---|---|
| GLM-5.2 | Z.ai | China | 51 |
| Muse Spark 1.1 | Meta | United States | 51 |
| MiniMax-M3 | MiniMax | China | 44 |
| DeepSeek V4 Pro | DeepSeek | China | 44 |
| Muse Spark | Meta | United States | 43 |
| MiMo-V2.5-Pro | Xiaomi | China | 42 |
| Kimi K2.7 Code | Moonshot AI | China | 42 |
| Hy3 | Tencent | China | 41 |
| DeepSeek V4 Flash | DeepSeek | China | 40 |
Has China caught the best closed U.S. AI models?
China has almost closed the overall quality gap, although the best closed American systems still perform better on the hardest frontier tasks.
On Artificial Analysis, Kimi K3 currently scores 57. That puts it close to GPT-5.6 Sol at 59 and Claude Fable 5 at 60. GLM-5.2 sits further back at 51, alongside several strong American models. The Stanford AI Index reached a similar broad conclusion earlier this year: the gap between the best American and Chinese models had fallen to low single digits and had changed direction several times.
The remaining difference becomes clearer on harder, narrower tasks. British and American government evaluators recently tested Kimi K3 on a 32-step simulated cyberattack. It reached step 17 on average, while the strongest American models reached 28.5. Kimi completed the full attack once in ten attempts. The average-progress gap was still large.
Kimi K3 also needs a small asterisk in an article about open models. Moonshot AI said the weights would become downloadable soon, but we could not verify an official public repository containing them during this review. For now, Kimi K3 proves that a Chinese lab can build a near-frontier model. Its contribution to the open-weight race begins once the files are actually available.
The gap is already narrow enough to change buying decisions. Many companies will accept a slightly weaker model when it is cheaper, customizable and available through several hosts. The top closed U.S. systems keep their edge on the hardest research, cyber, coding and agent tasks.
If you want more recent data on this point, please see our latest foundation model market report.
Is DeepSeek carrying China’s whole open-source AI industry?
China’s open-model lead no longer depends on DeepSeek, even though DeepSeek remains its biggest name.
GLM-5.2 currently shares the highest open-weight benchmark score. MiniMax-M3 matches DeepSeek V4 Pro on the same index. Xiaomi’s MiMo family has become one of the cheapest capable options, while Qwen covers more model sizes and modalities than almost any rival. Moonshot AI is pushing large agentic and multimodal systems, and Tencent has entered the leading group with Hy3.
The developer ecosystem shows the same spread. Hugging Face recorded sharp repository growth from Alibaba, Baidu, ByteDance, Tencent and several younger labs. MiniMax and Baidu both moved toward open releases after previously leaning more heavily on closed products. Stanford’s review of the Chinese ecosystem also found different commercial strategies and technical priorities rather than one national template copied by every company.
DeepSeek remains unusually influential. OpenRouter’s latest analysis found that V4 Flash became 70% of DeepSeek’s agentic token flow within roughly a month of release. That is unusually fast adoption for a new model used inside automated workflows.
The bigger advantage is choice. When one Chinese family disappoints, developers can move to another Chinese family without returning to a closed American API. That makes the lead harder to reverse than it looked during the first DeepSeek wave.
Are Chinese “open-source” AI models truly open source?
Most leading Chinese AI releases are open weight rather than fully open source.
The distinction is simple. Open weights let people download the trained parameters, run the model and often modify it. The Open Source Initiative asks for enough access to study and change the whole system, including the relevant code and detailed information about the training process and data.
Chinese labs usually publish the weights, model architecture, inference instructions and a license. They rarely release the complete pretraining dataset or everything needed to recreate the model from scratch. Western open-weight labs generally stop at the same point.
The licenses also differ. Qwen commonly uses Apache 2.0, while GLM-5.2 and DeepSeek V4 use MIT. Kimi’s modified MIT license requires prominent attribution once a commercial product passes very large user or revenue thresholds. MiniMax has used custom terms for some releases. A downloadable model can still carry commercial conditions.
For developers, open weights allow private deployment, fine-tuning, quantization and research beyond what a closed API permits. The accurate label is worth keeping. China is leading the open-weight market these days, while fully reproducible open-source frontier AI remains rare everywhere.
| Model family | Weights available | Typical license | What remains closed or restricted |
|---|---|---|---|
| Qwen | Yes | Apache 2.0 | Full training data and process |
| GLM-5.2 | Yes | MIT | Full training data and process |
| DeepSeek V4 | Yes | MIT | Full pretraining data |
| Kimi K2 family | Yes | Modified MIT | Attribution at very large commercial scale |
| MiniMax | Yes for major releases | MiniMax Community License for M3 | Commercial terms depend on the model; training disclosure remains incomplete |
Are developers really using Chinese open models, or only downloading them?
Developers are using Chinese open models for real work, and the adoption has become much harder to dismiss as launch hype.
Hugging Face downloads show reach, but they are an imperfect measure. A researcher may download several versions, and a popular model can be copied into hundreds of derivative repositories. OpenRouter gives us a better view of actual inference because it records the tokens processed through its platform.
The current weekly OpenRouter ranking is unusually Chinese. Tencent’s Hy3 leads with about 9.76 trillion tokens, followed by Xiaomi’s MiMo-V2.5 at 8.68 trillion and DeepSeek V4 Flash at 5.24 trillion. Three Chinese systems occupy the first three places on a platform used to compare models from around the world.
Automated workloads are helping them move quickly. Developers can compare price, speed and performance programmatically, then shift millions of tokens without asking end users to learn a new product. Chinese models are currently strongest in exactly that fluid part of the market.
OpenRouter’s audience is more price-sensitive and experimental than a typical bank or government department. Large enterprises also buy support, compliance and integration, areas where Anthropic, OpenAI and the American clouds remain stronger.
Even so, this story has moved well beyond GitHub stars and viral demos. Chinese models are processing large volumes of coding, agentic and general-purpose traffic. Adoption looks broad enough to persist even when individual model rankings change.
Is China winning mainly because its AI models are cheaper?
China’s low prices now matter because its open models have crossed the “good enough” quality line.
Artificial Analysis currently estimates a benchmark-task cost of roughly $0.04 for DeepSeek V4 Pro at a score of 44. MiniMax-M3 reaches the same score for about $0.12. GLM-5.2 scores 51 at roughly $0.47. GPT-5.6 Sol scores 59 at around $1.04, while Claude Fable 5 scores 60 at about $2.75.
DeepSeek will not create the same business value for one-fiftieth of Claude’s cost in every workload. A weaker model may need more attempts, more supervision or a second model to check its work. Enterprise support, uptime and security also cost money.
The order of magnitude is impossible to ignore. For a workload that already succeeds with DeepSeek or MiniMax, the premium model may cost ten to fifty times more per completed benchmark task. At high token volumes, a small quality improvement has to save a lot of human time before it pays for that difference.
Developers are already mixing models. An inexpensive Chinese system handles routine coding or agent steps, while a stronger closed model deals with the hardest cases. Chinese labs can win a large share of the workload underneath Claude and GPT without replacing either model everywhere.
| Model | Intelligence score | Estimated cost per task |
|---|---|---|
| DeepSeek V4 Pro | 44 | $0.04 |
| MiniMax-M3 | 44 | $0.12 |
| GLM-5.2 | 51 | $0.47 |
| GPT-5.6 Sol | 59 | $1.04 |
| Claude Fable 5 | 60 | $2.75 |
If you want more recent data on this point, please see our latest foundation model market report.
Can companies realistically run China’s best open models themselves?
The largest Chinese models require serious infrastructure, while smaller versions are practical for a much wider group.
The raw files make the problem visible. The GLM-5.2 repository occupies about 1.51 terabytes on Hugging Face. Kimi K2.6 is roughly 595 gigabytes. MiniMax-M3 has about 428 billion total parameters, with around 23 billion active for each token. GLM-5 has roughly 744 billion total parameters and 40 billion active.
Sparse mixture-of-experts designs reduce the computation used for each token, but the full weights still have to sit in memory or be distributed across several accelerators. Long context windows add another large memory bill. Running one of these systems reliably also requires serving software, monitoring, networking and people who understand model parallelism.
Most businesses will use a managed endpoint, a cloud deployment or a quantized version. Some will choose a smaller Qwen, MiMo or distilled DeepSeek model that fits on one or a few GPUs. They gain more choice over hosting and customization than they would with a closed API.
So “self-hostable” should not be read as “easy to run on an office server.” Companies mainly gain the freedom to choose their cloud, security setup and serving provider. Full independence is available, but it is expensive at the frontier.
Is China strong only in open-source language models?
China’s clearest lead is in language and coding, with a wider open ecosystem now spreading across vision, audio, images, video and agent tools.
Qwen is the best example of breadth. Its public repositories include general language models, coding systems, visual-language models, image generation, image editing, embeddings, speech recognition and text-to-speech. Moonshot AI has released multimodal Kimi models and an open audio foundation model. DeepSeek publishes OCR research and inference software, while MiniMax maintains tools spanning text, speech, images, video and music.
Comparisons become messier outside language models. We can test text systems on common coding and reasoning tasks. Image quality, video consistency, speech naturalness and embodied performance require different evaluations, many of which are controlled by the model creators.
The available data is too uneven to call China the leader in every open AI category, although the range of releases is already impressive. Chinese labs are building model families that can see, hear, generate media and use tools, rather than stopping at chat.
A company building an agent can now combine Chinese models for language, vision, speech and document understanding instead of waiting for one provider to supply everything.
Is China innovating in open AI or mainly copying U.S. models?
Chinese labs are doing original model engineering, even though they also learn aggressively from American research and model outputs.
DeepSeek has pushed sparse attention, mixture-of-experts routing and efficient reasoning systems. Moonshot AI trained Kimi K2, a one-trillion-parameter model with 32 billion active parameters, on 15.5 trillion tokens and developed MuonClip to prevent training instability. Qwen has built one of the widest families of model sizes and modalities. MiniMax has focused heavily on low active-parameter counts and long-context efficiency.
All of those teams build on ideas developed internationally. Mixture-of-experts models, reinforcement learning, synthetic data and distillation all came out of global research, and American labs also borrow architectures and techniques published elsewhere.
The disputed area is distillation: training a model partly from the outputs of a stronger model. U.S. companies and officials have accused some Chinese labs of collecting closed-model outputs at scale. Public information does not let us calculate the contribution to any current model, so the accusations remain one possible part of the story rather than a complete explanation.
Even perfect access to another model’s answers leaves the hard work of stable trillion-parameter training, efficient inference kernels, multimodal systems and a reliable release pipeline. China’s progress includes imitation, as frontier AI progress usually does, alongside real engineering capability.
Can China keep its open-model lead under U.S. chip restrictions?
China can keep producing competitive open models under the current restrictions, but the hardware gap makes a long-term lead harder to defend.
The imbalance is large. Stanford’s 2026 AI Index counts 5,427 data centers in the United States, more than ten times the number in any other country. It also estimates that U.S. private AI investment was 23 times China’s during 2025. Chinese government funds narrow the financial comparison, while America keeps a large advantage in advanced accelerators, cloud capacity and access to capital.
Chinese labs have compensated through sparse architectures, lower active-parameter counts, quantization and more careful use of compute. Huawei has also moved beyond prototypes. It recently showed a 1,024-card Atlas 950 SuperPoD and says more than 750 smaller Ascend 384 systems are already deployed.
The software remains the harder problem. One recent team completed full-parameter post-training of the DeepSeek V4 family on an Ascend SuperPOD, showing that serious model work is possible. A separate field study needed twelve source-level patches, disabled several high-throughput features and added safeguards against recurring device failures before DeepSeek V4 Flash ran reliably on sixteen Ascend devices. China has usable domestic compute, but engineers still pay a large migration tax compared with Nvidia’s mature CUDA stack.
Open weights help China work around part of the constraint. Once a model is released, American, European and Asian cloud providers can serve it on their own chips. The Chinese lab pays for training, while much of the global inference capacity comes from elsewhere.
Chip restrictions are slowing China and raising its costs. So far, they have encouraged efficiency without preventing Chinese labs from reaching the front of the open-weight market.
Will censorship and security concerns stop Chinese open models from spreading?
Security concerns will block Chinese models from some sensitive systems as broad commercial and developer adoption continues elsewhere.
Where the model runs changes the risk. Sending confidential data to a Chinese-operated API creates a different risk from downloading the weights and running them inside an American or European cloud account. In the second case, prompts can stay inside the customer’s chosen infrastructure.
The model itself deserves scrutiny. A recent U.S. government assessment found that GLM-5.2 allowed help with agentic cyber-exploit development and blocked fewer sensitive biological questions than the American comparison models. An independent study of Kimi K2.5 found narrow political censorship, especially in Chinese, along with relatively low refusal rates on some dangerous requests.
Open weights cut both ways. Researchers can inspect, test and modify the system more freely. A determined user can also remove safeguards. Origin-country bans become harder to enforce once the files are mirrored and served by third parties.
We expect a split market. Defense, critical infrastructure, government and heavily regulated companies will use stricter procurement rules or avoid some Chinese models entirely. Coding tools, startups, research projects and price-sensitive applications will keep adopting them when the performance and economics are attractive.
The security debate will reduce China’s reachable market. It is unlikely to reverse the wider diffusion of the weights.
If you want more recent data on this point, please see our latest foundation model market report.
Does China control the open AI ecosystem, or mainly the models?
China currently controls many of the most attractive open weights, while American companies control much of the infrastructure that turns those weights into products.
Amazon Bedrock now offers Qwen and several other Chinese model families as managed services. Microsoft Foundry supports DeepSeek. Google’s open-model platform includes DeepSeek and Qwen, and Nvidia has published serving support for recent Qwen models. Their presence across the major clouds shows real customer demand.
They also show where the money and control can remain American. The cloud provider handles deployment, identity, regional hosting, monitoring, scaling and billing. It owns the customer relationship and can swap one model for another inside the same platform.
Hugging Face, Nvidia’s software stack, vLLM and other serving tools play a similar role. Chinese labs influence what developers run, while foreign platforms often determine how those developers discover, optimize and pay for it.
Publishing weights lets Chinese companies reach users through foreign infrastructure. A closed Chinese API would have to fight for trust and distribution one customer at a time.
China has gained serious influence over the model layer, but it remains dependent on foreign platforms for much of its global distribution.
| AI layer | Current leader or advantage | Why |
|---|---|---|
| Frontier open-weight models | China | Most high-scoring downloadable models come from Chinese labs |
| Closed frontier models | United States | The strongest overall systems remain American |
| Global cloud distribution | United States | AWS, Microsoft and Google host Chinese models |
| Accelerator software | United States | Nvidia’s ecosystem remains the default for advanced deployment |
| Leading-edge chip fabrication | Taiwan | TSMC manufactures most frontier AI accelerators |
| Breadth of Chinese open releases | China | Several labs publish models across sizes and modalities |
Can the United States quickly take back the open-model lead?
One major U.S. release could retake first place, but matching China’s current depth would take longer.
Meta has already shown how quickly the ranking can move. Muse Spark 1.1 now matches GLM-5.2 on Artificial Analysis, which weakens any claim that China has an untouchable technical lead. Nvidia, Google, OpenAI and American startups also have enough talent and compute to release stronger weights.
The harder issue is incentive. OpenAI and Anthropic make money by selling controlled access to their best models. Publishing frontier weights would reduce that advantage and create safety, licensing and support problems. Meta and Nvidia have stronger reasons to promote open ecosystems because they can benefit from advertising, hardware or infrastructure around the models.
China’s position also includes volume. Developers can choose among DeepSeek, Qwen, GLM, MiniMax, MiMo, Kimi and Tencent models, often in several sizes. A single American benchmark winner would change the headline without immediately replacing that range.
We should expect the lead to move back and forth. The open-model market changes too quickly for a permanent national champion. Right now, though, the United States has the resources to catch up faster than China can close the chip and cloud gap.
So, is China winning open-source AI?
Yes, China is currently winning the open-weight model race, while the wider open-source AI stack remains divided.
As seen above, Chinese labs supply seven of the nine open-weight systems scoring 40 or above on the current Artificial Analysis table. OpenRouter has also recorded Chinese models overtaking American models in token volume for part of this year, based on more than 450 trillion observed tokens. Those two findings cover both capability and actual use.
China’s lead has depth. DeepSeek is important, but GLM, MiniMax, Qwen, MiMo, Kimi and Tencent give developers several alternatives. The models are usually cheap, commercially usable and available through major cloud platforms. They now compete for coding and agent workloads rather than attention alone.
The claim becomes too broad when people treat it as proof that China controls the whole AI industry. Most releases are open weight rather than fully open source. The strongest closed models remain American. U.S. companies dominate cloud distribution, accelerator software and private investment, while Taiwan remains central to advanced chip production. Security concerns will also keep some government and regulated customers away.
China leads the part of open-source AI that developers can download and build on today. The United States still owns more of the infrastructure, revenue and closed-model frontier around it. China is ahead in the model race, with the rest of the stack still out of reach.
If you want more recent data on this point, please see our latest foundation model market report.
OUR METHODOLOGY
This analysis tests whether China is winning open-source AI by separating the downloadable model race from the wider AI industry. We compare model capability, developer adoption, pricing, ecosystem depth, practical openness, infrastructure and distribution rather than treating one benchmark or launch as decisive.
For capability, we use current independent evaluations and live comparison tables, while keeping their limits in view. Benchmarks can contain errors, reward test-specific optimization or miss important differences in cyber, coding, agentic and multimodal performance, so we look for repeated strength across models and evaluators.
For adoption, we distinguish repository activity and downloads from actual inference. Hugging Face data shows the scale and breadth of Chinese releases, while OpenRouter token traffic gives a closer view of models being used in live workloads. OpenRouter is not global market share, but it is useful for tracking developers who actively compare price, speed and model quality.
We verify openness through official repositories, model cards and licenses. A model is treated as open weight when its trained parameters can be downloaded and deployed; we reserve “fully open source” for systems that also provide the broader code, training information and data access described by the Open Source Initiative.
We also keep the market layers separate. Leadership in downloadable weights does not automatically imply leadership in closed frontier models, cloud distribution, accelerator software, advanced chip fabrication, private investment or enterprise customer relationships.
Key sources used for this analysis include Hugging Face’s Spring 2026 review of open-model activity, OpenRouter’s large-scale usage study, OpenRouter’s DeepSeek V4 adoption analysis, OpenRouter’s current model rankings, Artificial Analysis on GLM-5.2, Artificial Analysis on Muse Spark 1.1, Stanford HAI’s 2026 AI Index, and the Open Source Initiative’s Open Source AI Definition.
We use official model repositories and cloud documentation to verify whether weights are public, which licenses apply, how large the models are and where they can be deployed. Important technical and distribution sources include Z.ai’s GLM-5.2 repository, DeepSeek’s public repositories, vLLM’s DeepSeek V4 serving documentation, Amazon Bedrock’s open-weight model announcement, and Microsoft Foundry’s managed-model documentation.
The conclusion comes from convergence. China is treated as ahead only where capability rankings, real usage, cost advantages, release depth and ecosystem adoption point in the same direction; claims about the broader AI stack are judged separately.
