- Current state of Chinese AI: leaders, market figures and commitment to open source.
- Comparison of the best platforms: DeepSeek, Doubao, ERNIE, Qwen, Kimi, Hailuo and KLING.
- Prices, appsreal strengths and limitations, with a focus on privacy and reliability.
If you're only familiar with ChatGPT , Gemini, Copilot , or Meta AI, it's time to look to the East: the Chinese artificial intelligence ecosystem is advancing at breakneck speed and is already competing head-to-head with Western giants. China has developed a catalog of foundational models, platforms, and tools that not only matches capabilities but also drives down costs and accelerates open-source innovation.
In this guide, we've condensed the best and latest from this ecosystem, from multimodal chatbots and reasoning LLMs to video generators and deep learning frameworks. We've compiled features, performance, apps, pricing, strengths and weaknesses , along with market context, so you can choose the platform that best suits your needs without getting lost in the noise.
Quick context: How China has gone from following in the footsteps to leading the way
The current boom began after the launch of ChatGPT in late 2022, when OpenAI demonstrated the conversational potential of LLMs. Google reacted late, Meta launched Llama , and the sector was shaken up.
By November 2024, the Cyberspace Administration of China had officially registered 309 major models , covering e-commerce, healthcare, education, finance, and cybersecurity . According to iiMedia, the Chinese market for models reached 29.416 billion yuan in 2024 and could exceed 70.000 billion yuan by 2026, with an average usage of 4,5 times per week compared to 2,1 in Western markets.
This sprint has a name: open source. Alibaba (Qwen) and others have released families of models that hundreds of thousands of developers are extending at a speed difficult for closed labs to match. The result is a vibrant ecosystem, with competitive models at reduced cost and real-world adoption at scale.
The best Chinese AI platforms (and what they're used for)
1) DeepSeek: powerful reasoning and affordable cost
DeepSeek made a splash with its V3 series and, above all, with R1 , a reinforcement-trained reasoning model that thinks "out loud" before responding. Its fame stemmed from combining performance and efficiency , with a reported training cost of $5,57 million, a fraction of that of other elite models.
- Advanced reasoning for complex problems.
- Good skills in programming (code and explanations).
- Fluency in mathematicswith step-by-step demonstrations.
With DeepSeek R1, you don't get an instant answer: the model first reflects (thinking chain), which brings transparency to demanding tasks. Users apply it to complex SQL, finance, and mathematical proofs . Version R1-0528 reduces hallucinations by around 45-50% in drafting, rewriting, or summarizing tasks.
The latest version from DeepSeek, V3.1-Terminus (September 2025), alternates between a fast mode for trivial queries and a deep reasoning mode for difficult problems. It also improves language quality and reduces Chinese-English mixing. In comparisons, ChatGPT's o3-mini can outperform it in mathematical accuracy , but with a significantly higher operating cost and lower performance in Chinese.
Availability and pricing: Web and apps (Android/iOS) with registration. Free to use in the general public version, with paid services via API. Also offers variants such as DeepSeek Math, Coder, and VL (vision).
2) Doubao (ByteDance): massive customization and a "warm" experience
Doubao is ByteDance's (TikTok/Douyin) big bet. Designed for a broad audience, it has conquered China with a more visual, engaging, and social experience . In 2024-2025, it amassed staggering figures: over 160 million registrations and peaks of 800.000 daily downloads ; in August, it reached 157 million monthly active users and was the most popular AI app in the country.
It works with text, voice, and video; it generates images, presentations, spreadsheets, podcasts, and 5-second clips. Its key differentiator is its intelligent agents with avatars, voices, names, and configurable personalities; you can even create your own, adjust privacy settings, and share it with the community. Its integration with Douyin redirects traffic and fosters virality.
In terms of performance, the Doubao 1.5-Pro series introduced enhanced inference capabilities and improved its rankings in both closed and open searches, competing with GPT-4 and Claude 3.5 Sonnet. Version 1.6 adds improved multimodality, significantly expanded context memory, and the ability to simulate click/drag actions, automating everyday workflows (e.g., exporting tickets to Excel).
Social trick: Doubao makes it easy to share results, and on Douyin, you can tag Doubao to summarize videos ; its responses recommend content that plays without leaving the app. It's where convenience and dopamine meet. Price: Free for users; business plan available.
3) ERNIE Bot (Baidu): language, multimodality and an emotional touch
Baidu , the dominant search engine in China, offers ERNIE Bot as a general-purpose assistant. Its latest iteration, ERNIE Bot 4.0 Turbo , generates multimodal content (text, images, websites, PDFs , and audio) and integrates with DeepSeek for in-depth searches with expert-level answers.
One interesting difference: its responses can be empathetic when you discuss personal topics. It excels at ideation, copywriting, analysis, and polishing long texts. It supports Chinese and English (with more limited English comprehension). Available on the web and as an app; registration is currently required.
Price: plans around $6,86/month , with official announcements of a completely free service starting April 1st (according to Baidu's own announcements). Furthermore, the group is boosting its ecosystem with PaddlePaddle (an open-source framework) and its own Kunlun chips.
4) Hailuo AI (MiniMax): Text to video with a focus on coherence and physics
Despite its relatively short history, MiniMax has made a strong impact with Hailuo . Its S2V-01-Live model brought text-to-video communication to massive audiences (over 27 million international views). The company boasts a prompt optimization feature that translates your intent into highly accurate instructions.
Version S2V-01 introduced "Subject Reference": you upload an image and the AI maintains the protagonist's visual identity while performing new actions, balancing fidelity and creative freedom without requiring extensive datasets or LoRA . It processes in seconds and drastically reduces computing costs.
With Hailuo 02, they took a leap forward in understanding the physical world: believable movements, realistic splashes , and better-executed complex scenes. It's like going from having a coherent actor to equipping them with a top-tier action director.
Comparison with Sora: Hailuo offers compelling and free trial pieces (1000 initial credits, 10-20 videos), while Sora offers a monthly subscription. It has a favorable cost-benefit ratio for experimenting. Price: Standard plan around $9/month.
5) Qwen (Alibaba): the open source “all-rounder”
Qwen is Alibaba Cloud's family of models and arguably the most influential in open source today. Thousands of derivatives circulate within the community, surpassing even Llama in the number of forks and variants. Qwen works with languages, multimodality, mathematics, and programming.
The Qwen 2.5-Max model shines in Chatbot Arena and excels in programming with its Artifacts feature , which renders web UIs in real time from code. Alibaba leverages its consumer insights for customizations, and its open strategy accelerates adoption and deployment.
Trained in 29 languages (German, Arabic, Spanish, among others), it offers good multilingual results. Available online with or without registration; official apps for Android/iOS, not always accessible in all countries. Price: free for general use, with associated cloud services.
6) Kimi (Moonshot): very long texts and visual reasoning
Moonshot has transformed Kimi into a favorite among professionals working with large documents. The platform became famous for handling up to 200.000 Chinese characters while maintaining consistency. In terms of traffic, Kimi even topped weekly rankings in China.
It's useful for information retrieval, document interpretation, academic writing with verifiable references , and spell checking. Its viral content generator is popular on networks like Xiaohongshu . On the downside, it can omit details or lose logic in very long texts and is slower with complex tasks.
The new K1 feature adds visual reasoning, demonstrating complete thought processes in image analysis (for example, explaining why a meme is effective). The experience is more engaging and can become a valuable educational tool. Price: Free version with a voluntary tip system.
7) KLING AI (Kuaishou): versatile generative video
KLING AI, from Kuaishou (a direct competitor of TikTok ), stands out for its variety of input options: text-to-video, image-to-video, video extension , and content filtering. The extension allows users to continue scenes from existing photos or clips while maintaining narrative and style.
Version 1.6 improves text responsiveness, dynamic quality, and visual aesthetics; in image-to-video, it boasts a 195% improvement over version 1.5, with more natural movements and expressions. It also offers virtual clothing try-ons for e-commerce.
Price: Free trials and a standard plan around $6,6/month . Its presence on Twitter with creative extensions has popularized the format.
Beyond the chatbot: frameworks, clouds and featured models
The Chinese tech industry's strength isn't limited to conversational assistants. Its technical and industrial capabilities are extensive , and it already competes globally across various frameworks and vertical markets.
- Baidu: PaddlePaddle (open source framework) and ERNIE; leadership also in autonomous driving with Apollo.
- AlibabaQwen and ModelScopeCity Brain for smart urban management.
- Tencent: Hunyuan for company; Tencent Miying deploys medical imaging diagnostics in more than 100 hospitals.
- Huawei: MindSpore optimized for Ascend chips; PanGu model with industrial and meteorological applications.
- iFlytekSpark LLM, strong in voice (ASR/TTS) and education.
- SenseTimeSenseNova/SenseChat; computer vision, smart cities and 3D avatars.
- Megvii: Face++ (vision APIs) and MegEngine (open source framework).
Startups to watch: Zhipu AI (GLM, a pioneer in agents and with plans to go public), Baichuan (Baichuan 2 open source), MiniMax (Hailuo) and DeepSeek (efficiency and advanced reasoning).
Foundational models visible on the radar: ERNIE 4.x , Qwen 2.5 , SparkDesk , PanGu 3.0 , SenseChat 5.0 , GLM-4 , Baichuan 2. They are the basis for multimodal solutions and applications by sector (health, education, retail, industry, public sector).
The Chinese “Big 5”: who sets the pace and why
A quintet leads the way: Alibaba, ByteDance, Stepfun, DeepSeek, and Zhipu . Each has a clear strategic focus: open source, mass-market consumer products, long-term research, or agents.
Alibaba has fully embraced Qwen and open source: releasing models and cultivating an ecosystem while monetizing cloud and premium services. It has announced massive investments in AI and infrastructure ( tens of billions of dollars) to accelerate adoption.
ByteDance brings what few others have: a massive global product and real-world engagement data. Their expertise with viral algorithms is being transferred to Doubao, which iterates with feedback from millions of users and leverages UX and distribution as a competitive advantage.
Stepfun maintains a low profile but aims high in benchmarks. Backed by state capital, it invests in patient research led by top talent (such as co-authors of ResNet), with strong results in OpenCompass and Chatbot Arena.
Zhipu , founded at Tsinghua University, promotes AI agents and concepts like Phone Use. It aims to be the first Chinese LLM startup to go public , which would publicly measure the value of these companies on markets in Hong Kong or Shanghai.
DeepSeek represents pure research as an advantage: teams focused on technical efficiency, without short-term commercial pressures , have shown that it is possible to compete with more limited budgets if the focus is on ingenuity.
Comparisons and practical notes that interest you
Performance: ChatGPT's o3-mini can outperform R1 in mathematical accuracy, but at a significantly higher operating cost and with worse performance in non-English-speaking languages such as Chinese. DeepSeek, on the other hand, boasts efficiency and transparency with explicit thought chains.
Generative video: Sora offers cinematic realism, but Hailuo provides free trials and one-click prompt optimization . Kling shines with its input modes and video extension over existing footage, useful for creators and brands.
Open vs. closed: Qwen's openness has triggered derivatives on Hugging Face and GitHub, while closed platforms like Ernie, Claude, and GPT are advancing with polished products and robust APIs. This mixed competition is accelerating innovation on a global scale.
Macroeconomic impact: The “democratization of computing” by Chinese AI is putting pressure on margins, influencing stock valuations, and restructuring production chains . Multimillion-dollar market drops have been seen on specific days, along with regulatory tensions (for example, regarding autonomous driving) due to national security concerns.
App availability, languages, and privacy
DeepSeek, Doubao, ERNIE, Qwen, Kimi, and KLING offer web access and, in most cases, Android/iOS apps (Qwen may not be available as an app in some countries). Doubao also has desktop versions (Windows and Mac ). ERNIE and Yuanbao (Tencent) require registration, although basic access is usually free.
Languages: Qwen is trained in 29 languages ; DeepSeek handles several well, including Spanish. ERNIE supports Chinese and English, with a better command of the former . Kimi and Doubao prioritize Chinese, although they offer increasingly global interfaces and content.
Privacy and reliability: As with any LLM, avoid sharing personal or sensitive data . Although hallucinations are less frequent, they still occur, and it's important to verify critical information. If you want to run models locally, check the official repositories on GitHub: DeepSeek, Kimi, and Qwen publish components and versions for private execution.
Passionate writer about the world of bytes and technology in general. I love sharing my knowledge through writing, and that's what I'll do on this blog, show you all the most interesting things about gadgets, software, hardware, tech trends, and more. My goal is to help you navigate the digital world in a simple and entertaining way.