Alibaba and Qwen 3: How competitive is China’s new AI?

Quick answers

• Which company makes Qwen? Alibaba. The models come from the Qwen team inside Alibaba Cloud, and Alibaba Group's CEO Eddie Wu leads the effort.

• Which country is Qwen from? China. Alibaba is headquartered in Hangzhou.

• Is Qwen open source? Partly. Many Qwen models ship as open weights, and the 27B model from August 2026 uses the Apache 2.0 licence. The largest flagship model carries a custom Alibaba licence.

• Can a European company use it? Yes, and self-hosting inside the EU is the cleanest way to keep your data under your own control. Which route fits you depends on your data, budget, and team.

If you want a straight answer for your own company, book a free AI consultation, and we will map the options in one session.

What Qwen actually is

Qwen (also written Tongyi Qianwen in China) is Alibaba's family of large language models. It covers chat, coding, image and video understanding, speech and agent tasks. Alibaba is better known as an online marketplace and cloud provider, but its cloud arm has spent years building and giving away models to pull developers onto its platform. For a longer look at that strategy, read our piece on Alibaba Cloud's global AI expansion.

The family moves fast. Here is where it stands, based on public reporting:

Dates for the Max weights differ between sources (12 or 17 August), so check the official Qwen pages on Hugging Face before you rely on any single date. For model-level detail, see our Qwen 3.8 Max article (currently in German).

Qwen vs ChatGPT, Claude and Gemini: what matters for a business

Benchmark charts are published by the vendors themselves, so treat them as a starting point. What changes your costs and risks is how you access the model.

Neither side wins everywhere. A managed product is typically quicker if only a general assistant is required to write emails for staff. An open model such as Qwen is a viable option if you need a private assistant who is more of a contract that, for whatever reason, doesn't want to leave your servers, or if you want to use someone that knows about your products, your contracts, or your tickets. In our guide to the best AI chatbots for business, we discuss the managed options.

Please read the file before building (License).

Not all open weights are created equal. The 27B model is licensed under Apache 2.0 (commercial use, modification, and distribution are permitted with the inclusion of licence notices). The larger Max version is licensed under Alibaba's own license. Qwen was released previously under the Tongyi Qianwen licence, which, according to the coverage, amounts to 100 million monthly active users and can only be used with permission of Alibaba.

In practice: open the LICENSE in the exact repository you want to deploy, save a copy of it, and have your attorney look it over before deployment. There is also no training data information provided with the publication of weights, and it is therefore not possible to check what the model has learnt. 

In simple terms, the genuine dangers of:

Content restrictions: Due to the fact that Qwen is from a Chinese company, responses concerning delicate matters can be limited or biased. This is generally unimportant for a customer support bot. It does for research or media purposes. In our article about the Chinese State's use of Chatbots to repeat state narratives, we demonstrate what to test for.

Where your data goes: If you use Alibaba's hosted API, the contents of your prompts will be processed by Alibaba. The models listed in Alibaba Cloud are different from those in the Frankfurt region; please check first before you sign up. It's not a worry if you're self-hosting.

Knowledge cut-off: The model has only the training information. Use it only as a link to your own documents; don't depend on it to tell you the news from last week.

EU rules: The EU AI Act establishes transparency requirements in its article on transparency, meaning that the user should be informed about the use of an AI system. Please look at the most recent application dates as there are suggested changes to the timeline.

Operations: Every few months Qwen releases a new generation, and a self-hosted model needs to be patched, monitored, and planed for its next generation. 

Three ways to use Qwen in your company

Our AI development team builds all three. For a no-code route that keeps data in the EU and supports open-source and on-premises models, we run CustomGPT AIQ, and you can see a working example in our chatbot case study. If you are comparing no-code tools first, start with our GDPR comparison for German SMEs. To test models on a laptop before spending anything, see Ollama vs LM Studio.

A 30-day pilot that settles the question

1. Week 1: pick one job. For example, answering internal policy questions or drafting replies to support tickets. One job, one owner.

2. Week 1–2: gather 100 real examples. Use actual questions and ideal answers from your team, not invented ones.

3. Week 2–3: run the same examples through Qwen and one other model. Score accuracy, tone, and refusals.

4. Week 4: Choose number-based decisions. Do a quality, cost per 1,000 answers, and hours your staff will need to manage it.

Do not skip step 4. The trick to the benchmark-heavy headlines is that it can be more about the operations than the licences.

What this means for German companies

Adoption is rapidly increasing. According to the statistics published by Destatis, 20% of German companies with 10 or more employees deployed AI tools in 2024, while 26% of the EU-27 companies did in 2025 (based on Eurostat statistics). Use is quite a bit greater in large firms (approximately 57% at 250+ employees) compared with firms with 10-49 employees (approximately 23%). That's the opportunity: a medium-sized company that has a well-governed, private assistant will be ahead of most of their competitors.  We go deeper on this in AI in German firms and on data location in EU on-premise vs US cloud.

Other open models worth testing alongside Qwen

A fair pilot compares at least two models. Good candidates include Kimi K3 and GLM 5.2, both covered on our blog.

Want Qwen, or any open model, running safely in your company?

Tell us your data, your team size and your use case. Our AI consulting team will tell you whether Qwen, another open model or a managed tool is the better fit, with a rough cost and timeline. Book your free consultation.

Frequently asked questions

Which company owns Qwen?

Alibaba. The models are built by the Qwen team within Alibaba Cloud.

Is Qwen from China?

Yes. Alibaba is based in Hangzhou, China, and Qwen is a Chinese model family.

Is Qwen free to use?

Some models are. Open-weight models such as Qwen3.8-27B can be downloaded and used commercially under Apache 2.0, but you pay for the hardware to run them. Alibaba's hosted API charges per token.

Is Qwen safe for business use in Europe?

It can be, if you control where it runs. Self-hosting or EU-hosted deployment keeps data under your control. Using a hosted API means a third party processes your prompts, so you need a data processing agreement and a clear view of data location.

Is Qwen better than ChatGPT?

It depends on the task. Vendor benchmarks show strong results in coding and reasoning, but ChatGPT, Claude and Gemini remain easier to adopt as ready-made products. Test both on your own examples.

Can I run Qwen on my own servers?

Yes, for open-weight models. Alibaba positions the 27B model for local and on-premise use. Hardware needs depend on the model size and how it is compressed, so size your setup with a pilot.

Which Qwen model should I start with?

For most company pilots, a mid-sized open model such as the 27B is the practical starting point. The largest models need far more infrastructure.

Sources

• Alibaba Qwen models and weights: https://huggingface.co/Qwen and https://qwen.ai/

• eWeek, Alibaba Qwen3.8-27B licence coverage: https://www.eweek.com/news/alibaba-qwen3-8-27b-license-apac-china/

• DataNorth, Qwen3.8-27B release details: https://datanorth.ai/news/alibaba-releases-qwen3-8-27b

• Eurostat, use of AI in enterprises (isoc_eb_ai): https://ec.europa.eu/eurostat/databrowser/view/isoc_eb_ai/default/table

• EU AI Act, Article 50: https://artificialintelligenceact.eu/article/50/

Editor's notes (delete before publishing)

• Keep the same URL (/en/blogs/alibaba-qwen-3-how-competitive-is-chinas-new-ai/) and replace the content, so the existing impressions carry over. Do not publish this as a new URL.

• Overlap: /blog/qwen-3-8-max/ already covers the model news in German and has no English twin in your sheet. This post is the business and decision page and links to it. Create the English twin and swap the link. Do the same for the German version of this post (/blog/alibaba-qwen-3-wie-wettbewerbsfahig-ist-chinas-neue-ki/) with an hreflang pair.

• Search queries this targets: which company, which country, is Qwen from Alibaba, is Qwen Chinese, Qwen 3 test report. Keep the Quick answers block at the top, because it matches those queries directly.

• Fix on the live page: change author from Tech-Now-Client to a named person, replace the generic affiliate-marketing CTA with the one above, update the Similar Posts to AI model and open-source posts, and set a visible updated date.

• Schema: Article, FAQPage (the seven FAQs), Organization, Person.

• Verify before publishing: Max weights date and licence terms, the 100 million MAU clause in the legacy licence, Alibaba Cloud Frankfurt model availability, the 2025 German figures at Eurostat or Destatis, and the EU AI Act dates.

• Linked pages marked PENDING in your sheet: service pages, case study and most blog links. Confirm they are live and indexable first.

• City rule: no German city is named, so no city data was added.