Alibaba: Qwen Models, Cloud and Checkpoint Licenses

Alibaba Group develops Qwen language and multimodal models and runs cloud infrastructure alongside its commerce businesses. Its published checkpoints have different licenses, including Apache 2.0 and model-specific terms.

kindCompany
founded1999

Latest

Updated

Models within a cloud and commerce group

Alibaba was established in 1999 by 18 founders led by Jack Ma. The group now combines commerce with cloud infrastructure and AI services. It develops the Qwen family of language and multimodal models and uses that technology across enterprise services and consumer platforms.

Qwen3 brought two mixture-of-experts and six dense models to developers in April 2025. The release supplied Apache 2.0 weights and supported both thinking and non-thinking modes, giving developers a choice of model sizes and inference behavior.

Licensing remains checkpoint-specific. Qwen3.8-27B-FP8 lists Apache 2.0, whereas Qwen3.8-2.4T-A95B uses the Qwen3.8-Max License. The latter includes a separate-license condition for qualifying Model-as-a-Service or AI Work Assistant businesses above its stated revenue threshold. Public weights do not imply identical commercial terms across the family.

In August 2026, Alibaba priced a HK$80 billion share placement and said it intended to put all net proceeds into its AI capabilities and infrastructure. The announcement described a conditional expected closing, rather than establishing that the proceeds had already been spent on capacity.

Sources

  1. [1]
    Introduction to Alibaba Group

    Alibaba Group · retrieved Oct 24, 2026

  2. [2]
  3. [3]
    Qwen/Qwen3.8-2.4T-A95B

    Hugging Face · Aug 12, 2026

  4. [4]
    Qwen3.8-Max License

    Qwen · retrieved Oct 24, 2026

  5. [5]
    Qwen/Qwen3.8-27B-FP8

    Qwen · retrieved Oct 24, 2026

  6. [6]

Alibaba history

2026

│
Qwen-Image-2.1 ships open weights under a research-only license

One model for text-to-image and editing, with a 7 billion parameter image generator, native transparent layers and up to 10 reference images. The weights are public, but the license covers research and evaluation only, where the original Qwen-Image was Apache 2.0.

Qwen
Model
│
Anthropic accuses seven Chinese labs of distilling Claude

With Anthropic, Moonshot AI, DeepSeek, Xiaomi

Anthropic's September misuse report named Alibaba, Moonshot, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax, and said Moonshot, DeepSeek and Xiaomi fed their own users' conversations into Claude. It also described a blocked request for help with gain-of-function work on chikungunya virus.

Anthropic
Research
│
Alibaba launches Wan3.0, a video model that reads slide decks

Wan3.0 doubles Alibaba's clip length to 30 seconds and takes PDFs, spreadsheets and slide decks as input alongside the usual text and images, aimed at corporate marketing teams as much as filmmakers. It ships a day after Alibaba priced a $10.2 billion Hong Kong share sale earmarked for AI infrastructure, and unlike its predecessor Wan2.7, the weights stay closed.

Dataconomy
Model
│
Qwen 3.8 27B lands under Apache 2

A 27 billion parameter vision-capable model from Alibaba, permissively licensed and small enough to run on a workstation. The most capable model most people could actually own kept being made in China.

Hugging Face
Model
│
Alibaba publishes 2.4 trillion parameters

Qwen3.8 Max went up on Hugging Face under Alibaba's own license: 2.4 trillion parameters with 95 billion active, 262,000 tokens of context natively, text only and thinking always on. Second largest set of open weights in the world, sixteen days after the largest.

Hugging Face
Open source
■
Anthropic accuses Alibaba of its largest distillation attack yet

With Anthropic

A letter to the Senate Banking Committee alleged 28.8 million exchanges with Claude through 25,000 fraudulent accounts between April and June, aimed at reproducing Claude's coding and agentic capabilities in Alibaba's own Qwen models. Alibaba did not respond publicly.

CNBC
Company

2025

│
Qwen3 ships eight open models with a thinking switch

Two mixture of experts models and six dense ones from 0.6B to 32B, all Apache 2.0, all able to turn step-by-step reasoning on and off per request. Hybrid thinking was the pitch; the licence was the reason the family spread.

Qwen
Model
│
Alibaba commits $53 billion to AI and cloud infrastructure

A three-year spending plan Alibaba said exceeds its total AI and cloud investment over the prior decade, announced alongside six straight quarters of triple-digit growth in AI product revenue. CEO Eddie Wu named artificial general intelligence the company's primary long-term goal.

Alibaba Cloud
Deal
│
Alibaba launches Qwen2.5-Max, its flagship model, against DeepSeek

Alibaba released Qwen2.5-Max, a mixture-of-experts model pretrained on more than 20 trillion tokens, and published benchmark comparisons claiming wins over DeepSeek V3 on Arena-Hard, LiveBench and GPQA-Diamond. It shipped through Qwen Chat and Alibaba Cloud's API rather than as open weights.

Qwen
Model

2024