Which Chinese AI lab is ahead?







Every funding round, every investor, and what the labs promised in return, in one document. Have a look ↓
Summary
Alibaba's Qwen is ahead overall among Chinese AI labs, with the strongest combination of competitive models, a massive open-weight developer ecosystem, and an established business capable of turning AI adoption into revenue. DeepSeek remains its most important challenger, while Xiaomi currently leads on some model intelligence benchmarks.
The technical race is much closer than the commercial one. Xiaomi's MiMo-V2.6-Pro scores around 46 on Artificial Analysis's Intelligence Index, while the leading models from Alibaba, DeepSeek and Z.ai sit around 45. Such narrow differences make it difficult for any laboratory to maintain a convincing performance lead. Differences of a single point are why we separate who leads on benchmarks from who only claims it in our report on frontier AI labs.
Qwen's strongest advantage comes from what developers have already built around it. Hugging Face identified more than 151,000 derivative repositories based on Qwen, creating an ecosystem of adaptations, tools and deployment experience that competitors cannot reproduce simply by releasing a better model.
DeepSeek has a different kind of momentum. Its share of OpenRouter token traffic doubled from approximately 9% to 18% during the first half of 2026, with strong adoption of V4 Flash for agentic workflows. Developers are using its models heavily, not merely downloading them.
Xiaomi is the clearest reminder that China's AI hierarchy can change quickly. Its leading open-weight model currently outranks Qwen's proprietary flagship on one composite benchmark, despite Xiaomi lacking anything close to Alibaba's established AI developer ecosystem.
Inference economics are also becoming more complicated than advertised token prices. DeepSeek V4 Flash has appeared at extremely low API prices, yet Xiaomi's MiMo costs roughly half as much per task on one standardized Artificial Analysis workload. The cheapest model per token is not necessarily the cheapest model for a completed task.
Moonshot AI and Z.ai remain serious technical challengers. Their leading Kimi and GLM models are close to China's frontier, especially for coding and reasoning, but neither has built the breadth of smaller models and developer adaptations that distinguishes Qwen.
The biggest commercial contrast is between Alibaba and the independent laboratories. Alibaba generated around $7.1 billion in quarterly cloud and compute revenue, while MiniMax recorded $116.6 million across the first half of 2026. The comparison is imperfect because Alibaba sells much more than models, but it reveals an enormous difference in business scale.
MiniMax shows how quickly that gap can start to narrow in specialized markets. Its first-half revenue rose 283%, largely driven by enterprise and platform services, although adjusted losses of $293 million underline the expense of competing at the frontier.
Baidu remains an important player despite receiving less attention for frontier benchmarks. Its quarterly AI-powered businesses generated RMB12.5 billion, giving it significant commercial reach even as DeepSeek, Qwen, Xiaomi and other specialists attract more interest for their latest models.
American proprietary models still lead the broadest intelligence rankings, but Chinese laboratories have become especially competitive in open-weight AI. Alibaba is best positioned to benefit from that shift because Qwen combines strong model quality with distribution, developer loyalty and cloud commercialization. A rival could take the benchmark lead tomorrow without taking Qwen's overall lead.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us
70+pages
This is just a preview page. Don’t worry if a number looks old: the report gets refreshed, and you’ll always get the latest version.
Why is it so hard to name China's leading AI lab?
China has several world-class AI labs competing closely, and no single company leads in model intelligence, developer adoption, actual usage and revenue at the same time.
DeepSeek changed the international perception of Chinese AI with its inexpensive reasoning models. Alibaba's Qwen built a much larger ecosystem of downloadable models and developer adaptations. Xiaomi has since produced one of China's highest-scoring open-weight models, while Moonshot AI, Z.ai and MiniMax are competing seriously at the frontier.
The technical differences have become surprisingly small. Artificial Analysis currently puts Xiaomi's MiMo-V2.6-Pro at roughly 46 on its Intelligence Index, followed closely by Alibaba's Qwen3.8 Max, DeepSeek's V4.1 Flash 1017 and Z.ai's GLM-5.3, all around 45. These are close results for models built by companies with very different resources and business strategies. Scores this close from very different labs make ranking hard, so we say who is ahead in our frontier AI labs report.
Yet the commercial picture looks different. Hugging Face's August 2026 analysis found that Qwen had created a much larger open-weight ecosystem than its Chinese competitors. OpenRouter's first-half 2026 research, meanwhile, showed DeepSeek doubling its share of token traffic.
For an overall winner, we should care about which laboratory has built a position that can last beyond the next model release. On that basis, Alibaba starts with an advantage, although DeepSeek and Xiaomi have strong reasons to challenge it.
70+pages
Just a preview. If a number here looks old, no worries: your report comes with the latest data.
Is DeepSeek still China's most important AI lab?
DeepSeek remains one of China's most important AI labs, especially for cheap and widely used inference, but it no longer has a clear technological lead over its Chinese competitors.
DeepSeek became internationally famous after releasing DeepSeek-V3 and DeepSeek-R1, challenging the assumption that frontier reasoning models required the budgets of the largest American AI companies. Its April 2026 V4 release and subsequent updates showed that the company could build on that breakthrough.
More importantly, developers continued using its models. OpenRouter examined more than 450 trillion tokens processed between January and mid-June 2026 and found that DeepSeek's share of platform traffic had climbed from around 9% to 18%. It had briefly approached 20% in early June, after falling as low as approximately 5% earlier in the year.
The recovery came alongside changes in how people were using DeepSeek. By late May, V4 Flash accounted for around 70% of the company's agentic token traffic on OpenRouter. Developers were clearly using it heavily in automated workflows, rather than merely testing a newly released chatbot.
Still, the performance gap has narrowed. Xiaomi now scores higher on some composite intelligence evaluations, Alibaba has the larger developer ecosystem, and Moonshot and Z.ai remain credible alternatives for demanding reasoning tasks.
DeepSeek deserves its reputation. But China's AI competition has moved beyond the period when its name dominated almost every international discussion.
Which Chinese AI lab has the smartest model right now?
Xiaomi currently has the highest-scoring Chinese open-weight model on Artificial Analysis's Intelligence Index, although Alibaba, DeepSeek and Z.ai are close behind.
MiMo-V2.6-Pro scores approximately 46, compared with around 45 for Alibaba's Qwen3.8 Max, DeepSeek's V4.1 Flash 1017 and Z.ai's GLM-5.3. Moonshot's Kimi K3 and StepFun's Step 5 Preview follow closely.
Artificial Analysis evaluates models across coding, scientific reasoning, knowledge work and automated tasks. Its composite index allows us to compare models using a common methodology rather than relying on benchmarks selected by their creators.
The small gaps are important. A difference of one or two points on this index does not necessarily mean that a model will produce better answers for every customer. DeepSeek, for example, outperforms Xiaomi on certain automation and professional-work evaluations despite scoring slightly lower overall.
Xiaomi's results are particularly impressive because the company is primarily known for smartphones, vehicles and consumer electronics. It also has a potential distribution advantage through those products, although we have not yet seen evidence that MiMo commands a developer ecosystem comparable to Qwen's.
Another distinction concerns availability. Alibaba's Qwen3.8 Max is a hosted proprietary model, while Xiaomi's leading MiMo release has downloadable weights. Alibaba offers separate open-weight models, including Qwen3.8 2.4T A95B, which currently scores around 40.
| Chinese AI lab | Leading model | Intelligence Index | Availability |
|---|---|---|---|
| Xiaomi | MiMo-V2.6-Pro | ~46 | Open weights |
| Alibaba | Qwen3.8 Max | ~45 | Proprietary |
| DeepSeek | V4.1 Flash 1017 | ~45 | Open weights |
| Z.ai | GLM-5.3 | ~45 | Open weights |
| StepFun | Step 5 Preview | ~44 | Model-dependent |
| Moonshot AI | Kimi K3 | ~44 | Open weights |
| MiniMax | MiniMax-M3.1-Flash | ~43 | Open weights |
| Alibaba | Qwen3.8 2.4T A95B | ~40 | Open weights |
These are rounded Artificial Analysis scores for specific model versions. Rankings can change when new models or benchmark updates arrive.
70+pages
This is a preview page. Spotted an outdated number? That’s normal here. The report you get is always the freshest version.
Has Alibaba's Qwen overtaken DeepSeek?
Alibaba's Qwen has already overtaken DeepSeek in the breadth of its developer ecosystem, although DeepSeek remains exceptionally strong in API usage.
Hugging Face's August 2026 research recorded approximately 2.045 billion downloads for Qwen repositories with declared parameter counts during its observation period. Moonshot recorded around 37 million under the same measurement, giving Qwen roughly 55 times as many downloads. Open-weight models that developers download by the billion are a technology gaining traction, and we follow it in our report covering frontier AI labs.
The broader comparison is just as striking. Hugging Face identified 151,448 derivative repositories based on Qwen, approximately 2.6 times Meta's total derivative footprint and 4.7 times the repositories specifically derived from Llama.
These derivative repositories include models adapted for languages, industries, specialized tasks and different hardware. Each successful adaptation gives other developers another reason to start with Qwen.
DeepSeek's position looks stronger when we examine inference traffic. Its rapid growth on OpenRouter suggests that developers are choosing it for applications that repeatedly call an AI model, sometimes generating enormous numbers of tokens.
Both laboratories can legitimately claim success. Qwen is more deeply embedded in the open-weight development community, while DeepSeek has demonstrated particularly strong demand for running models.
For long-term influence, we give Alibaba the advantage. Qwen's ecosystem is already so much larger that a modest benchmark lead from a rival probably won't be enough to overturn it.
Why are so many AI developers choosing Alibaba's Qwen?
Developers keep choosing Qwen because Alibaba offers a wide range of model sizes, flexible licenses and tools that make it relatively easy to build specialized AI applications.
Hugging Face found that Qwen derivatives were appearing at a rate of approximately 180–210 repositories per day during the first seven months of 2026. At the midpoint of that range, this amounts to roughly 41,500 additional repositories over 213 days.
That's a lot more than developers briefly downloading a fashionable model. Thousands are taking Qwen as a starting point for their own work.
Alibaba supports this by releasing everything from compact models suitable for local deployment to much larger systems intended for powerful servers. Qwen3.8-27B, released in August 2026, gives developers another option when they want to run a capable model without paying for access to an enormous hosted system.
The contrast with some competitors is revealing. Hugging Face found that 88% of Moonshot's measured downloads came from models larger than 70 billion parameters, compared with 55% for DeepSeek. MiniMax's measured downloads were also concentrated in very large models.
Smaller models are useful for organizations with limited computing budgets, specialized applications and businesses that need to keep their data on their own hardware. Alibaba's frequent use of Apache 2.0 licensing for smaller Qwen releases makes commercial adaptation easier as well.
Qwen has consequently become useful to a much wider group of developers than a laboratory focused almost exclusively on giant models.
70+pages
Preview only. Some numbers may have changed since we made this page, but your report will be up to date.
Is Moonshot AI's Kimi still competitive with Qwen and DeepSeek?
Moonshot AI's Kimi remains competitive with China's best AI models, especially on difficult reasoning and coding tasks, but its developer ecosystem is much smaller than Alibaba's.
Kimi K3 currently scores around 44 on Artificial Analysis's Intelligence Index, close to the best Chinese models. Kimi K3 Max has also appeared among the leading Chinese entries in Arena's human-preference evaluations, including coding comparisons.
This continues a strategy Moonshot has followed through successive Kimi releases: building powerful models for demanding work, rather than trying to offer the widest possible collection of smaller models.
That gives Moonshot a clear audience. Developers working on complex software projects or research problems may value the capabilities of a large Kimi model more than access to dozens of smaller alternatives.
Z.ai is pursuing a similar opportunity with its GLM family. The progression from GLM-4.5 through GLM-5 and subsequent updates shows repeated investment in coding, reasoning and agentic workflows. GLM-5.3 now sits within roughly one point of Alibaba's leading hosted model on Artificial Analysis.
Neither laboratory has yet matched Qwen's broad development ecosystem. But both are close enough in model performance to compete for users who care more about getting a difficult task done than about which company has the most downloads.
Is DeepSeek still the cheapest Chinese AI model worth using?
DeepSeek is still among the most attractive low-cost Chinese AI providers, particularly for applications processing enormous volumes of tokens, although Xiaomi and other competitors now challenge it on cost efficiency.
OpenRouter reported pricing for DeepSeek V4 Flash as low as $0.09 per million input tokens and $0.18 per million output tokens on its cheapest available endpoint in June 2026.
At those prices, an application processing 100 million input tokens and generating 20 million output tokens would pay approximately $12.60. At $1 per million input tokens and $3 per million output tokens, the same workload would cost $160. Ten-fold price gaps for the same job show how differently labs earn, and we compare which business models are winning in our frontier AI labs market report.
That's nearly thirteen times as much for the same token volumes, assuming no differences in caching, discounts or the amount of reasoning needed.
But Xiaomi's MiMo-V2.6-Pro shows why cheap token prices are only part of the calculation. Artificial Analysis reports a cost of approximately $0.13 per task for MiMo on its standardized Intelligence Index workload, compared with $0.26 for DeepSeek V4.1 Flash 1017 in the same comparison.
MiMo is therefore cheaper on that particular benchmark workload, despite DeepSeek's appeal as a low-cost API. DeepSeek's model also runs considerably faster in the comparison, which can matter for interactive applications.
The right choice depends on what the application does. DeepSeek is especially compelling for heavy, repetitive inference, while Xiaomi may offer better economics for some difficult reasoning tasks. Picking whichever provider advertises the lowest token price would be a mistake.
Which Chinese AI lab is getting the most real usage?
DeepSeek has some of the strongest independently measured usage among Chinese AI labs, but nobody has enough public data to identify the country's biggest model provider across every platform.
OpenRouter's first-half 2026 research provides unusually useful evidence because it measures tokens actually processed by developers. As seen above, DeepSeek doubled its share of platform traffic over the period and became OpenRouter's leading model provider by token volume from mid-May.
That suggests substantial usage in production applications, especially when combined with the high proportion of traffic generated by V4 Flash in agentic workflows.
However, a large share of Chinese AI usage happens elsewhere. Alibaba customers can use Qwen through Alibaba Cloud. Baidu sells enterprise AI services through its own infrastructure. Other organizations run downloaded models on private servers, sometimes without reporting their usage publicly.
These workloads are mostly invisible to OpenRouter's data.
Token consumption also needs interpretation. A model that produces longer answers or uses more reasoning tokens can generate greater traffic without necessarily serving more customers. A low-priced model can process enormous volumes while earning relatively little revenue.
DeepSeek has demonstrated strong demand on one of the most visible independent AI marketplaces. That makes it a credible usage leader, but a nationwide or worldwide ranking would require data that the industry does not currently disclose.
70+pages
This is just a preview page. Don’t worry if a number looks old: the report gets refreshed, and you’ll always get the latest version.
Is Alibaba making more money from AI than DeepSeek and its other rivals?
Alibaba has a much larger disclosed AI-related cloud business than China's independent model startups, giving it a commercial advantage that DeepSeek, Moonshot and Z.ai have not matched publicly.
Alibaba reported approximately $7.1 billion in AI Cloud and Compute Services revenue for the June 2026 quarter, up 45% year over year. AI-related product revenue had also grown at triple-digit rates for twelve consecutive quarters. Growth like Alibaba's AI cloud makes it fair to ask who the top companies are today; we answer that in our report on frontier AI labs.
This reflects a long-running commercial shift rather than a sudden burst of interest following one model launch. The cloud business also reported adjusted EBITA of approximately $830 million, with a margin near 12%.
For comparison, MiniMax generated $116.6 million in total revenue over the entire first half of 2026. Alibaba's reported cloud and compute revenue for just one quarter was therefore roughly 61 times MiniMax's six-month company revenue.
That comparison has an important limitation: Alibaba's figure includes cloud infrastructure and computing services beyond Qwen itself. It would be misleading to describe the entire amount as model revenue.
Baidu offers another useful benchmark. Its AI Cloud Infrastructure business generated RMB7.3 billion in the June quarter, approximately $1 billion at prevailing exchange rates, with 50% year-over-year growth.
Alibaba's overall commercial scale is clearly ahead of these disclosed figures. What remains less transparent is how much revenue comes directly from each laboratory's models, rather than from selling the infrastructure and services surrounding them.
Is MiniMax growing faster than China's bigger AI labs?
MiniMax is growing remarkably fast as a specialized Chinese AI company, with first-half 2026 revenue almost quadrupling, although it remains much smaller and less profitable than Alibaba's cloud business.
MiniMax's published financial results show revenue increasing from $30.4 million in the first half of 2025 to $116.6 million in the first half of 2026, a rise of 283%.
The interesting part is where that growth came from.
Open Platform and other enterprise AI services jumped from $9.2 million to $73.9 million, an increase of 703%. Consumer-oriented AI-native products also grew strongly, from $21.2 million to $42.6 million, but the enterprise and platform business expanded much faster.
We can calculate that approximately $64.7 million of MiniMax's $86.2 million additional revenue came from enterprise and platform services. That's about 75% of the company's total year-over-year growth.
The change is visible in the revenue mix. Enterprise and platform services accounted for 30% of revenue a year earlier and 63% in the latest six-month period.
MiniMax also reported that token consumption had grown twentyfold between January and July 2026, giving us another indication that paid model usage is becoming central to its business.
Profitability is still a serious problem. The company recorded an adjusted net loss of $293 million for the first half, around 2.5 times its revenue, despite improving its gross margin from 12.1% to 17.9%.
MiniMax has proved that it can attract paying customers quickly. The next test is whether its revenue can keep growing without requiring losses several times larger than sales.
70+pages
Just a preview. If a number here looks old, no worries: your report comes with the latest data.
Is Baidu falling behind DeepSeek and Alibaba in Chinese AI?
Baidu has fallen behind the most visible Chinese frontier-model labs in technical attention, but its AI business is still growing quickly and already generates billions of renminbi in quarterly revenue.
Baidu's June 2026 results showed AI Cloud Infrastructure revenue reaching RMB7.3 billion, up 50% year over year. AI applications generated another RMB2.5 billion, while AI-native marketing services contributed RMB2.6 billion.
Together, these three categories reached RMB12.5 billion and accounted for approximately half of Baidu's General Business revenue. A year earlier, their combined share had been 38%.
Baidu is becoming more dependent on AI-related businesses as traditional search advertising faces pressure.
The company also has substantial existing distribution. Baidu's main app reported 644 million monthly active users in June 2026, while ERNIE Assistant had reached 202 million monthly active users in December 2025.
Those numbers do not establish that Baidu's latest language models outperform DeepSeek or Qwen. They do show that Baidu has ways to introduce AI products to hundreds of millions of people and sell related services to businesses.
We would put Baidu behind the strongest frontier-model specialists on current technical momentum, but ahead of many startups in revenue, enterprise infrastructure and consumer reach.
Can DeepSeek and Moonshot really compete with Alibaba's money?
DeepSeek, Moonshot and other independent Chinese labs can compete with Alibaba on model performance, but matching Alibaba's computing infrastructure and distribution is much harder.
Alibaba enters the AI race with an existing cloud business, enterprise customers, semiconductor and infrastructure partnerships, and cash-generating operations outside its model laboratory. That makes it easier to fund expensive model development across several generations.
The pressure on smaller laboratories is clear from MiniMax's financial statements. The company held approximately $1.323 billion in cash at the end of June 2026, but its first-half adjusted net loss reached $293 million. It has substantial resources, yet the cost of competing remains high. A lab burning cash with $1.3 billion in the bank still needs backers; we look at who is getting funded, and by whom, in our frontier AI labs report.
DeepSeek has a different starting point through its association with High-Flyer, while Moonshot and Z.ai have relied on external investment and revenue from their own products and services.
The strong benchmark results of Xiaomi, Z.ai and DeepSeek show that capital does not translate neatly into model intelligence. Smaller teams can still produce excellent systems through architecture choices, data, training efficiency and focused engineering.
Alibaba's advantage becomes more pronounced after those models are ready. It can distribute them through its cloud, sell computing services around them and offer integrated products to customers it already serves.
Independent labs can absolutely beat Alibaba in individual model releases. Competing with its entire AI business requires much more than winning a benchmark.
Have Chinese AI labs caught up with OpenAI, Anthropic and Google?
Chinese AI labs are now serious competitors to the best American models, especially in open-weight AI, but the leading US proprietary systems still perform better on broad intelligence evaluations.
Artificial Analysis currently scores Anthropic's Claude Opus 5.5 at roughly 58 and OpenAI's GPT-6 Astra at approximately 53. Xiaomi's leading MiMo model sits around 46.
The gap between Claude and MiMo is therefore approximately twelve index points. That is still substantial on a common evaluation system, despite the rapid progress made by Chinese developers.
China's position looks stronger when we compare models that developers can download and run themselves. Xiaomi, DeepSeek, Alibaba, Moonshot and Z.ai have all contributed powerful open-weight systems, giving developers alternatives to paid proprietary APIs.
Hugging Face's research also found that several Chinese laboratories concentrated their releases in very large open-weight models. Moonshot, MiniMax and Z.ai have helped make this part of the market much more competitive.
Developers can now access sophisticated Chinese models, customize them and deploy them on infrastructure they control.
For the moment, the best American proprietary models still lead on general intelligence scores. Chinese developers, meanwhile, have become essential competitors in the market for powerful, customizable and relatively inexpensive AI.
70+pages
This is a preview page. Spotted an outdated number? That’s normal here. The report you get is always the freshest version.
Which Chinese AI lab has the strongest overall position?
Alibaba's Qwen currently has the strongest overall position in Chinese AI, combining competitive model quality, enormous developer adoption and access to a profitable, fast-growing cloud operation.
We can see why by comparing the actual advantages of its competitors.
Xiaomi has an edge on recent open-weight intelligence evaluations. DeepSeek has attracted heavy usage for inexpensive inference. Moonshot and Z.ai are near the frontier for reasoning and coding. MiniMax has demonstrated unusually fast growth in paid enterprise services, while Baidu already serves a large consumer and commercial market.
Alibaba is the only contender among them with such a strong combination of model competitiveness, open-weight reach and cloud commercialization.
Its developer ecosystem also gives Qwen a potentially lasting advantage. Every successful adaptation adds to the pool of tools, specialized models and deployment experience available to other developers.
This advantage could weaken if Qwen's model quality fell far enough behind the competition. For now, however, its open-weight ecosystem remains much larger than those of its nearest Chinese challengers.
| Competitive question | Strongest Chinese contender | Supporting evidence | Main limitation |
|---|---|---|---|
| Best open-weight model | Xiaomi | MiMo-V2.6-Pro scores around 46 | Small benchmark lead |
| Broadest developer ecosystem | Alibaba Qwen | More than 151,000 derivative repositories | Downloads do not equal paid usage |
| Strongest measured OpenRouter usage | DeepSeek | Leading token provider from mid-May 2026 | Only one platform measured |
| Largest disclosed AI cloud business | Alibaba | Around $7.1B in quarterly cloud and compute revenue | Includes infrastructure beyond Qwen |
| Fastest-growing measured pure-play business | MiniMax | Revenue up 283% year over year | Heavy ongoing losses |
| Strong specialist reasoning rival | Z.ai | GLM-5.3 scores around 45 | Smaller development ecosystem |
| Strong large-model alternative | Moonshot AI | Kimi K3 scores around 44 | More concentrated model portfolio |
| Established AI consumer and cloud reach | Baidu | RMB12.5B in quarterly AI-powered business revenue | Weaker frontier-model position |
Alibaba leads across more of the areas that could keep it competitive for years, while its rivals have narrower but sometimes impressive advantages.
Which Chinese AI lab is ahead overall?
Alibaba's Qwen is currently China's leading AI lab overall, although DeepSeek remains its strongest challenger in low-cost inference and Xiaomi has taken the lead on some open-weight model benchmarks.
Our research points to Alibaba because it combines several advantages that competitors have only achieved separately. Qwen has one of the world's largest ecosystems of model adaptations, its flagship models remain close to China's technical frontier, and Alibaba has a major cloud business capable of selling AI infrastructure and services to existing customers.
DeepSeek has a different strength. Its ability to attract intensive developer usage shows that inexpensive, capable models can win substantial demand even against companies with much larger distribution networks.
Xiaomi's rapid progress is another reason the ranking could change. Its MiMo models demonstrate that frontier AI development is no longer confined to traditional AI laboratories, and the company's consumer hardware business creates opportunities to distribute its technology widely.
Z.ai and Moonshot are close enough in model quality to remain serious competitors, while MiniMax's revenue figures show that specialized Chinese AI businesses can attract paying customers at impressive speed.
For Alibaba to lose its overall lead, one of these challengers would probably need to combine technical success with a much larger developer or commercial ecosystem. A single higher benchmark score would not be enough.
As of now, Qwen is the clearest overall winner. But China's strongest individual model, most heavily used low-cost API and broadest AI developer ecosystem belong to different competitors. That is why the race remains open, even with Alibaba ahead.
70+pages
Preview only. Some numbers may have changed since we made this page, but your report will be up to date.
Our methodology
This analysis compares China's leading AI laboratories across four dimensions: model intelligence, open-weight developer adoption, actual model usage and commercial performance. We assess Alibaba Qwen, DeepSeek, Xiaomi, Moonshot AI, Z.ai, MiniMax and Baidu using independent evaluations, platform activity and company financial disclosures rather than choosing a winner based on one benchmark.
For technical performance, we use Artificial Analysis's Intelligence Index and its benchmarking methodology to compare specific model versions under a common evaluation framework. We also consider Arena's human-preference leaderboard for additional context on model quality. Intelligence scores are rounded, and small differences should not be interpreted as decisive advantages across every workload.
Developer adoption is assessed primarily through Hugging Face's State of Open Models, Summer 2026. We examine model downloads, derivative repositories, model-size distribution and the growth of adaptation activity. Download counts and derivative repositories measure different kinds of engagement, and neither directly establishes the number of paying customers.
For actual inference usage, we rely on OpenRouter's DeepSeek V4 adoption research, which examines more than 450 trillion tokens processed during the first half of 2026. We use its findings to understand DeepSeek's share of platform traffic and its adoption in agentic workflows. OpenRouter represents one visible distribution channel, so we do not treat its traffic rankings as a complete picture of Chinese or global AI consumption.
Our commercial comparison draws on Alibaba's June-quarter financial disclosure, Baidu's second-quarter results and MiniMax's first-half 2026 financial results. We compare reported revenue, growth, profitability and business mix while distinguishing model-related revenue from broader cloud infrastructure and AI-powered services. These figures come from different business structures and reporting periods, so they are used to assess scale rather than to construct a direct model-revenue ranking.
We also use the official Qwen3.8-27B model card, Qwen's model-family announcement, Z.ai's model repositories and MiniMax's model card to examine model availability, licensing and release strategies.
Our overall judgment favors laboratories with advantages that extend beyond a single model generation. We weigh technical competitiveness alongside evidence of recurring developer adoption, substantial usage, business scale and distribution. This is a qualitative assessment based on the combined evidence, not a mechanically weighted ranking. On that basis, Alibaba's Qwen has the strongest overall position, with DeepSeek, Xiaomi and the other leading Chinese laboratories retaining meaningful advantages in individual categories.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us









