Alibaba 阿里巴巴

Company profile

Founded: 1999
Headquarters: Hangzhou
Main AI products:
- Qwen 通义千问 series:
  - Qwen 3.5 (released 2026-02-16 on Lunar New Year’s Eve): 397B total / 170B active parameters, natively multimodal; Alibaba claims performance at or above Gemini 3 Pro and GPT-5.2; API input price at ¥0.8 per million tokens (about one-eighteenth of Gemini 3 Pro)
  - Qwen 3 series (released 2025-04)
  - Qwen 2.5 / 2 / 1 series
- Wan 万相: video generation; Wan 2.6 released 2025-12
- Tongyi Lab 通义实验室: Alibaba’s AI R&D hub
- Bailian 百炼 (Model Studio): enterprise-facing LLM API platform
- Quark 夸克: AI browser / assistant
Business model: e-commerce + cloud (Alibaba Cloud + Tongyi) + payments (Ant Group, separate) + logistics + AI applications

Across-the-board Apache 2.0 open source: Qwen has 400+ open-sourced models, 200,000+ derivatives, and over 1 billion downloads
32% market share of China’s enterprise LLM API market, #1 (2026 Q1 data)
Benchmarked against Meta Llama: open weights + commercialisation services in combination
Alibaba Cloud + Tongyi: dual-engine, with AI as Alibaba Cloud’s second growth curve

Type	Document	Link	Subpage
Usage policy	Qwen service agreement + open-source licence	tongyi.aliyun.com	—
Technical reports	Qwen 3 / 3.5 Technical Report	Qwen site + arXiv	—
Open-source models	Qwen GitHub / HuggingFace	github.com/QwenLM	—
Transparency	CAC algorithm filings (multiple batches)	cac.gov.cn	—

China:
- Algorithm filing under the Generative AI Interim Measures — Qwen was among the first-wave filed services
- TC260-003-2024 compliance
- Alibaba participates as a leading company in the drafting of multiple TC260 standards
United States: Qwen open-source models can be freely downloaded; Quark and other consumer products are not offered in the U.S.
European Union: Qwen open-source models are downloadable in the EU; AI Act Article 53 obligations (training-data summary) would apply if triggered
Southeast Asia / Middle East: active overseas push (Qwen-for-SEA and other localised variants)

Alibaba Qwen is the most successful open-source AI ecosystem from China:

Metrics (disclosed as of April 2026):

400+ open-source models (Qwen 1/2/2.5/3/3.5 series + Wan series + multimodal variants)
200,000+ derivative models (fine-tuned on top of Qwen)
Over 1 billion downloads (HuggingFace + ModelScope combined)
32% market share of China’s enterprise LLM API market, #1

Commercial logic:

Open source as customer acquisition: open weights → developer experimentation → conversion to paid Alibaba Cloud API (analogous to Meta Llama + AWS)
Tongyi + Alibaba Cloud dual engine: AI as Alibaba Cloud’s second growth curve
Aggressive pricing: Qwen 3.5 API at ¥0.8 per million input tokens — about one-eighteenth of Gemini 3 Pro

Qwen’s AI Act compliance position in the EU:

Qwen 3.5 training compute very likely exceeds 10^25 FLOP → triggers systemic-risk GPAI
The open-source exemption in AI Act Art. 53(2) does not apply to systemic-risk GPAI
In theory, Qwen must meet: adversarial testing, systemic-risk assessment, incident reporting, and cybersecurity
In practice: Alibaba has not signed the GPAI Code of Practice (as of 2026-04-22)
Whether the EU AI Office will enforce against scenarios where Qwen open weights are downloaded into the EU is an open boundary

Qwen’s position in the U.S.:

Open-source models are freely downloadable in the U.S.
Alibaba Cloud International operates in the U.S. but does not include the Qwen main line (U.S. compliance risk)
Potentially affected by future BIS export-control policy targeting “Chinese AI weights”

Alibaba is one of the global #1 open-source AI providers, yet it has not published a standalone safety-framework document:

No RSP / Preparedness / FSF equivalent
No standalone AI Principles (only a sub-section embedded in the Alibaba ESG report)
Not a participant in the Frontier Model Forum
No public red-team disclosure mechanism

Explanations:

Regulator interaction is primary: governance obligations discharged via CAC filings and TC260 participation
Open source is itself transparency: technical reports, model weights, and training details are public
Commercial pragmatism: similar to Meta / Mistral, avoiding “commitments that constrain future flexibility”

Dimension	Alibaba Qwen	Meta Llama	Mistral
Flagship open source	Qwen 3.5 (Apache 2.0)	Llama 3/4 (Llama Community License, use restrictions)	Mistral Large 3 (MRL, research-only)
Open-source depth	Deepest (including flagship)	Medium (commercial with conditions)	Medium (flagship closed)
Self-regulation document	None	Responsible Use Guide (weak)	None formal
Regulatory engagement	China: strong; elsewhere: weak	U.S.: medium; EU: signed CoP with partial reservations	EU: strong; signed CoP
Business model	Alibaba Cloud subscription	Advertising + Meta products	API + Mistral Compute

Commercial logic of the open-source strategy: open source for acquisition + cloud services for monetisation; closer to Meta + AWS than to Anthropic
No standalone safety framework: governance obligations discharged via domestic compliance and technical reports
Alignment with national strategy: TC260 standard participation, support for national AI industrial policy
Industrial-AI orientation: B2B and vertical applications (finance, healthcare, manufacturing) over consumer AI
International compliance remains to be observed: unsigned GPAI CoP; potential BIS impact

Dimension	Alibaba Qwen	Meta Llama	Mistral
Flagship parameters	397B total / 170B active	~405B	675B / 41B active
License	Apache 2.0	Llama 3 Community License (use restrictions)	Apache 2.0 (partial)
Home jurisdiction	China	US	France (EU)
Business model	Cloud + subscription	Advertising + Meta products	API + Mistral Compute
Number of open-source models	400+	~15	~20