For many people outside China, “Chinese AI” still means one name: DeepSeek.
That view is already outdated.
Kimi now has one of China’s strongest frontier models and a surprisingly complete AI work environment. Alibaba’s Qwen combines a genuinely capable free assistant with one of the industry’s largest model ecosystems. Z.ai is pushing hard into coding and long-running agents. MiniMax is combining coding with image, speech, music and multimodal AI.
DeepSeek still matters enormously — but not necessarily for the reason most people think.
So instead of asking which Chinese chatbot gives the best answer to ten prompts, we compared the actual products: intelligence, coding, research, documents, agents, multimodal tools, pricing, open weights and how practical they are for someone outside China.

The quick verdict
If you want one Chinese AI for the widest range of serious work, our current pick is Kimi.
If you want to spend nothing, start with Qwen Studio.
If you are building with AI and care about low API costs, self-hosting and permissive licensing, DeepSeek remains exceptionally compelling.
If coding and long-running technical agents dominate your work, Z.ai deserves serious attention.
And if you want coding alongside image, voice, music and broader multimodal creation, MiniMax is arguably the most interesting bundle.
| Priority | Our pick |
|---|---|
| Best overall AI workspace | Kimi |
| Best free general AI | Qwen |
| Best API/value proposition | DeepSeek |
| Raw frontier intelligence | Kimi K3 ≈ GLM-5.3 |
| Best measured Chinese coding-agent setup | Kimi Code + K3 |
| Technical/coding specialist | Z.ai |
| Creative & multimodal bundle | MiniMax |
| Broadest open-model ecosystem | Qwen |
| Strong self-hosting proposition | DeepSeek |

DeepSeek is no longer China’s intelligence leader
This is probably the most important correction to the old narrative.
On Artificial Analysis’ current Intelligence Index, Kimi K3 at maximum effort scores 60, tied with GLM-5.3 max at 60. Qwen3.8 Max scores 58, while the latest DeepSeek V4 Pro scores 53. MiniMax M3 sits at 45, and Tencent Hy3 at 42.
That does not mean Kimi will beat DeepSeek on every question. Composite benchmarks combine different kinds of reasoning, coding, knowledge and agentic tasks, and model settings matter.
But it does mean that buying DeepSeek because you assume it remains China’s smartest AI is no longer supported by the evidence.

Kimi is the surprise overall winner
Kimi wins for a reason that has little to do with winning every benchmark:
it has become unusually complete.
The normal Kimi web and mobile product handles chat, search, Deep Research, websites, presentations, documents and spreadsheets. Kimi Work can work with local files and automate desktop workflows. Kimi Code operates in the terminal and IDE. Kimi Claw is designed for persistent cloud automation.
That makes Kimi much closer to an AI operating environment than a conventional chatbot.
Its entry Andante membership costs ¥49 per month and includes Agent usage, office-file processing, Deep Research, website deployment, scheduled tasks, plugins and project storage. Higher tiers add more parallelism, Agent Swarm, Goal Mode, Kimi Claw and larger allowances.
There is a catch: Kimi’s paid features share a common credit pool. A demanding Deep Research task or agent workflow therefore reduces the capacity available elsewhere. Kimi Code also has a separate five-hour-per-week limit, and access to the newest K3 coding model can depend on membership level.
So Kimi is not “unlimited ChatGPT for ¥49.”
It wins because, among the Chinese products we examined, it currently creates the fewest situations where you need to leave the ecosystem to finish the job.

Qwen is the one to try before paying anyone
Qwen has perhaps the simplest value proposition in this entire comparison:
Qwen Studio is free to use and open to all.
And this is not a stripped-down demo.
Qwen Studio includes reasoning, web search, multimodal understanding, Deep Research, image generation, video generation and web-development tools.
Its current flagship, Qwen3.8 Max, scores 58 on Artificial Analysis and supports a context window of up to one million tokens along with image and video inputs.
The wider ecosystem is equally important. Alibaba has now released Qwen3.8 model weights, including the enormous 2.4-trillion-parameter family and much more practical smaller variants such as Qwen3.8-27B. The hosted Max product adds capabilities such as built-in tools and default million-token context.
Alibaba is also building QwenWork, a workplace agent platform aimed at autonomous professional work.
The problem for a global reader is that QwenWork’s public beta currently remains China-only.
Until that changes, Kimi has the stronger all-round international product. But if your first question is simply:
“Which strong Chinese AI can I try without another subscription?”
Qwen is the obvious starting point.

DeepSeek’s killer feature is economics
DeepSeek has lost the uncontested benchmark crown, but it may still have the most disruptive economics.
DeepSeek V4 Pro provides a one-million-token context window, is available through the web, app and API, and has downloadable open weights.
Its official API platform currently lists V4 Pro at:
- $0.435 per million cache-miss input tokens
- $0.87 per million output tokens
- just $0.003625 per million cache-hit input tokens
Prices can change, but those figures are extraordinary for a model scoring 53 on the current intelligence benchmark.
More importantly, DeepSeek V4 Pro is released under the MIT licence. That gives developers unusually broad freedom to use, modify, redistribute and commercially deploy the software and weights.
This means DeepSeek can lose the consumer-product comparison and still be the better decision for a company processing millions of tokens.
For developers, the relevant question may not be:
“Is DeepSeek smarter than Kimi?”
It may be:
“How much will this workflow cost when it runs 50,000 times?”
On that question, DeepSeek remains extremely difficult to ignore.

Coding is where Kimi and Z.ai become especially interesting
Chinese models are also much closer to the Western coding frontier than the old stereotype suggests.
The August GLSRM Coding Agent benchmark ranks Kimi Code CLI + Kimi K3 at 61.3 on its Coding Agent Index. That places it below the strongest Claude Code and Codex configurations, but comfortably inside the frontier group. A Claude Code configuration using Qwen3.8 Max scored 58.7, while Codex running DeepSeek V4 Flash scored 55.5.
The important word here is configuration.
GLSRM measures the complete agent — model, tools, harness and context-management system — rather than pretending the model alone determines coding performance.
That is why we would not simply declare “Kimi is the best coding model.”
And Z.ai complicates the picture further.
Its new GLM-5.3 ties Kimi K3 at 60 on Artificial Analysis, while Z.ai reports major improvements in complex coding and long-horizon tasks over GLM-5.2.
Z.ai’s Coding Plan starts at a promotional $12.60 per month, versus an $18 standard monthly price, with 10,000 weekly credits and support for more than 20 coding/agent tools.
GLM-5.3 is simply too new to have the same depth of independent full-agent evidence as Kimi K3.
So the current verdict is:
Kimi has the strongest independently measured Chinese coding-agent configuration we found. Z.ai is the specialist most capable of overturning that result.

MiniMax proves that raw IQ is not everything
MiniMax M3 scores only 45 on the same intelligence index.
That number badly understates why MiniMax is interesting.
M3 has a one-million-token context window, native image and video understanding, coding capability, agentic tools and computer use.
Then look at the subscription.
MiniMax’s $20/month Plus Token Plan includes roughly 1.7 billion M3 tokens per month, with text, image, speech and music sharing the same usage pool. Higher tiers raise the token allowance substantially.
MiniMax has also continued expanding its creative stack, including Music 3.0 and its H3 multimodal generation family.
That makes MiniMax less like “another ChatGPT” and more like a developer and creative-AI toolkit.
If all you need is the smartest text model, choose something else.
If you need coding, agents and several creative modalities from one ecosystem, MiniMax suddenly makes much more sense.

Open weights are a major Chinese advantage — but read the licence
This is one area where the Chinese market can be genuinely more attractive than the traditional Western subscription model.
DeepSeek V4 Pro uses MIT.
Tencent Hy3 uses commercially permissive Apache 2.0.
Qwen has now released enormous Qwen3.8 weights alongside more deployment-friendly smaller versions.
Kimi K3 is also downloadable, but it has its own Kimi K3 licence rather than MIT or Apache.
MiniMax M3 similarly uses a custom community licence with commercial-use conditions.
So “open-weight” should never be translated automatically into:
“I can do anything I want commercially.”
The model licence is part of the product.

Tencent is the dark horse
Tencent did not make our headline five, but it came much closer than expected.
Hy3 scores 42 on Artificial Analysis, so it is not currently competing for the raw-intelligence crown. But it is fast, inexpensive and Apache 2.0 licensed.
More importantly, Tencent has pushed it internationally through WorkBuddy, Miora and Tencent Cloud TokenHub. WorkBuddy can handle research, data analysis, documents, presentations and multi-agent workflows, while Hy3 is also available to developers through APIs.
That combination — global availability, permissive licensing and very low inference cost — makes Tencent one of the contenders most worth watching.
The global-user caveat matters
A benchmark can tell you how capable a model is.
It cannot tell you whether its privacy terms, payment system, regional access or regulatory environment fit your work.
DeepSeek, for example, states in its current privacy policy that personal data collected through its consumer services is directly collected, processed and stored in the People’s Republic of China.
That is not evidence that DeepSeek is unsafe, and it would be wrong to assume every Chinese provider has identical policies.
It does mean that confidential business data, client files, regulated information and sensitive research should never be uploaded merely because the model is cheap.
Read the actual provider’s current privacy and commercial terms first.
Can these replace ChatGPT, Claude, Gemini or Perplexity?
In some workflows, absolutely.
Kimi is the closest Chinese challenge to ChatGPT’s broad AI-workspace strategy.
Kimi and Z.ai are serious competitors for Claude-style coding and difficult knowledge work.
Qwen and MiniMax challenge parts of Gemini’s multimodal proposition, although neither reproduces Google’s enormous Gmail, Drive, Docs and YouTube ecosystem advantage.
Kimi and Qwen offer serious research tooling, although Perplexity still has a particularly clean source-first research experience.
But perhaps the most disruptive competition is happening somewhere else entirely:
price and control.
Western AI products have generally competed by building increasingly polished subscription ecosystems.
Chinese AI companies are increasingly offering a different combination:
frontier capability + low API prices + downloadable weights + agentic products.
For businesses building their own systems, that can matter more than which chatbot has the nicest interface.
Which one should you choose?
Choose Kimi if:
You want one Chinese AI environment that can move between research, documents, coding, local files and agentic work.
Our overall pick.
Choose Qwen if:
You want an excellent general AI without paying first, or you want access to Alibaba’s rapidly expanding open-model ecosystem.
Our free pick.
Choose DeepSeek if:
You care about APIs, self-hosting, low inference costs and a permissively licensed flagship model.
Our developer/value pick.
Choose Z.ai if:
Coding and long-running technical agents are your main workload.
Our specialist pick.
Choose MiniMax if:
You want coding and agents alongside image, speech, music and broader multimodal creation.
Our creative/developer pick.

Final verdict
The most important finding from this comparison is not that another Chinese company has “beaten DeepSeek.”
It is that there is no longer one Chinese AI story.
Kimi is building a full AI workspace.
Qwen is making a remarkably broad AI product available for free while releasing increasingly powerful model weights.
DeepSeek is attacking the economics of AI deployment.
Z.ai is specialising in serious coding and long-running agents.
MiniMax is combining technical AI with a much wider creative stack.
And Tencent is emerging internationally with a low-cost, permissively licensed dark horse.
If we had to choose one Chinese AI for the broadest range of work today, we would choose Kimi.
If we did not want to pay, we would start with Qwen.
If we were building a high-volume AI system and cared more about economics and control than the consumer interface, we would look extremely closely at DeepSeek.
That is the bigger shift.
Chinese AI is no longer interesting because it is simply cheaper than Western AI. It is interesting because its strongest companies are now competing on a different combination of intelligence, openness, agents, multimodality and cost.


























