Subscribe
China
7 min read
axc
axc

What Is Qwen? Alibaba's AI Models Explained

Qwen is Alibaba's AI model family. The Qwen3.8 lineup, how its licenses differ, how to try it, and what Alibaba has said about Qwen 4.

Olivetti-style cover reading WHAT IS QWEN with three stacked color blocks for model sizes in red, mustard and teal on cream paper
Olivetti-style cover reading WHAT IS QWEN with three stacked color blocks for model sizes in red, mustard and teal on cream paper

Qwen is the family of AI models built by Alibaba. If you have seen a model name like Qwen3.8-Max, Qwen3.8-27B or Qwen-Image-2.1, it comes from the same team. Some Qwen models are offered through Alibaba's cloud API. Others have weights you can download and run on your own hardware.

This guide explains what Qwen is, what is in the current lineup, how the licenses differ, how to try the models, and what Alibaba has said about Qwen 4. Facts are as of November 4, 2026. For how Qwen fits next to DeepSeek, Kimi, GLM and other Chinese labs, see our guide to China's AI labs.

Who makes Qwen

Qwen models are published under the Qwen organization on Hugging Face and the QwenLM organization on GitHub. The research license on one Qwen model names the legal entity behind it. The Qwen-Image-2.1 license defines "We" as "Hangzhou Tongyi Laboratory Technology Co., Ltd."

Alibaba presents Qwen as part of a larger plan. In its September 22, 2026 Apsara Conference press release, Alibaba described updates "featuring Qwen foundation models and multimodal models, new proprietary AI chips, a purpose-built agentic cloud, and an AI agent platform for mobile phones."

The current lineup

Qwen is not one model. It is several product lines that share a name. The main text models today belong to the Qwen3.8 generation. The diagram below shows the three Qwen3.8 text models with published weights, using the sizes and licenses from each model card.

Diagram of the Qwen3.8 lineup: Qwen3.8-2.4T-A95B with 2.4T total and 95B activated parameters under the Qwen3.8-Max License, Qwen3.8-27B with 27B parameters under Apache 2.0, and Qwen3.8-Flash-Next with 125B and 6B activated under the Qwen Community License 1.0

Qwen3.8-Max and its open weights

Qwen3.8-Max is the flagship. Alibaba published the open weights behind it as Qwen3.8-2.4T-A95B, uploaded to Hugging Face on August 8, 2026 (UTC). The model card lists "2.4T in total and 95B activated" parameters.

The model card says: "For the first time, Qwen3.8 brings a Qwen-Max-class model to open release." It also says the hosted version is different from the download: "Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools, etc."

The Qwen Cloud page for Qwen3.8-Max describes it as a "2.4-trillion-parameter MoE flagship" and lists image, text and video as inputs and text as output.

Qwen3.8-27B

Qwen3.8-27B is the compact dense model, uploaded on August 5, 2026 (UTC). Its model card calls it "a native vision-language model that understands images and videos, with flexible thinking control." The card lists 27B parameters.

Qwen3.8-Flash-Next

Qwen3.8-Flash-Next was uploaded on August 24, 2026 (UTC). Its model card lists "125B with 6B activated, plus 51B n-gram embedding and 4B MTP" parameters. The card calls it an "experimental preview of the architecture that will underpin Qwen4." It also says the hosted Qwen3.8-Flash is "the official version based on Qwen3.8-Flash-Next with more production features."

The card names four changes in this architecture: hybrid attention with Qwen Sparse Attention (QSA), a "Gated Residual" design, n-gram embeddings, and a training recipe that applies the Muon and AdamW optimizers to different weight categories.

Image models

Qwen also makes image models. The main image model on the Qwen Hugging Face page is Qwen-Image-2.1, uploaded on September 14, 2026 (UTC). Its card describes "a unified text-to-image generation and image editing model" with "7B parameters in its visual generation component." The card says it can make transparent (RGBA) images and supports "up to 10 reference images" for editing. A faster variant, Qwen-Image-2.1-Turbo, followed on October 9, 2026 (UTC). Alibaba's September 22 press release says Qwen-Image 3.1 is "set to launch later this year."

Open weights do not all mean the same license

Each Qwen model has its own license, and they are not the same. Always read the LICENSE file in the repository you download.

  • Qwen3.8-27B: Apache 2.0. The model card metadata lists license: apache-2.0.
  • Qwen3.8-Flash-Next: Qwen Community License 1.0. The LICENSE file has two conditions. The first says that if the model is used for commercial products or services "that have more than 100,000,000 monthly active users or US$ 20,000,000 (or equivalent in other currencies) monthly revenue, respective model name must be prominently displayed on the user interface of such product or service." The second says: "If the licensee or any of its affiliates conducts a Model as a Service or AI Work Assistant business, the licensee shall obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose." It adds: "The foregoing requirement shall not apply to the licensee's internal Use of the Software, provided that such Use does not make the Software, its outputs, or its underlying model capabilities available to any third party."
  • Qwen3.8-2.4T-A95B: Qwen3.8-Max License. The LICENSE file has the same first condition. Its second condition adds a revenue trigger: "If the licensee or any of its affiliates conducts a Model as a Service or AI Work Assistant business, and the aggregate revenue of the licensee and its affiliates exceeds US$50,000,000 (or the equivalent amount in any other currencies) during any consecutive twelve (12) months, the licensee shall obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose." The same internal-use exception follows.
  • Qwen-Image-2.1: Qwen Research License Agreement. The LICENSE file is a separate agreement dated September 20, 2026. It grants rights "for non-commercial purposes only" and says commercial use needs a separate license from Qwen.

Both the Community License and the Qwen3.8-Max License define the two business types. "Model as a Service" means "giving a third party access to language model inference or fine-tuning (e.g., via API or a hosted endpoint) in a manner that allows such third parties to exercise meaningful control over the inputs, parameters, or training data." "AI Work Assistant" means "an independent AI-powered product primarily designed for AI-assisted coding or office productivity (e.g., Qoder and QwenWork)."

QwenWork is the English name of Alibaba's Qwen Office (千问办公) desktop agent. Our report on Qwen Office 2.0 covers its November 2026 update, which adds team spaces and event-triggered tasks.

This guide does not give legal advice. If you plan to build a product on a Qwen model, have a lawyer read the license for that exact model.

How to try Qwen

There are three main ways to use Qwen, depending on what you want.

Diagram of three ways to use Qwen: chat in Qwen Studio, call the Qwen Cloud API, or download weights from Hugging Face
  1. Chat in a browser. The Qwen3.8-2.4T-A95B model card has a Qwen Studio button that links to Qwen's chat site, chat.qwen.ai. The Qwen Hugging Face page calls the same site Qwen Chat.
  2. Call the API. The model cards say: "the official Qwen API service is provided by Qwen Cloud." Each model has its own page, such as Qwen3.8-Flash, with features, rate limits and pricing. The Qwen3.8-Max page lists features including function calling, structured outputs, batches, web search and fine-tuning. Prices change, so check the model page before you build.
  3. Download the weights. Open-weight models are on Hugging Face. The model cards say the files work with Hugging Face Transformers, vLLM, SGLang and TokenSpeed (the 2.4T card lists vLLM, SGLang and TokenSpeed).

Note that the hosted and downloaded versions are not always the same. For both Max and Flash, the model cards say the hosted "official version" adds features such as a 1M context length by default and built-in tools.

Qwen 4: what Alibaba has said

Alibaba's September 22, 2026 press release says: "At the conference, Alibaba revealed that its next-generation model, Qwen 4, is currently in training." It adds: "The company further announced its roadmap for the upcoming Qwen 4.5 and Qwen 5 model series, projected to scale up to 5 to 10 trillion parameters."

The same release reports a self-improvement test. Alibaba says that over a month of fully automated runs, Qwen3.8-Max "completed 33 iterative cycles," and that "the updated Qwen3.8-Max boosted its Artificial Analysis score from 40 to 45." Those are Alibaba's figures, published September 22, 2026.

On October 29, 2026, an attendee at Alibaba's Apsara event in Amsterdam posted that Qwen 4 would come "next week." We covered that report in Qwen4 Next Week? What Alibaba Said in Amsterdam. As of November 4, 2026, the Qwen organization on Hugging Face listed no Qwen 4 model. The newest upload was Qwen-Image-2.1-Turbo, from October 9, 2026 (UTC).

Why Qwen matters

Alibaba offers its Qwen flagship as a hosted API and has also published downloadable weights at several sizes, from a 27B dense model to the 2.4T-parameter model that Qwen3.8-Max is based on. Users can call the API, or run the weights under the license for that model.

A model being downloadable does not mean every use is free of conditions. Apache 2.0 on the 27B model and the Community License on Flash-Next have different rules for the same kind of business. For a closer look at the Qwen3.8 release itself, see our Qwen3.8 explainer.

Quick answers

Is Qwen open source? Some Qwen models have downloadable weights, but the licenses differ. Qwen3.8-27B uses Apache 2.0. Qwen3.8-Flash-Next uses the Qwen Community License 1.0, and the 2.4T model uses the Qwen3.8-Max License, both with the conditions quoted above. Qwen-Image-2.1 uses the Qwen Research License Agreement, which is for non-commercial use only.

Is the downloaded Qwen3.8-Max the same as the API version? No. The model card says the hosted Qwen3.8-Max adds "vision input & non-thinking support, 1M context length by default, official built-in tools, etc." over the downloadable Qwen3.8-2.4T-A95B.

Can Qwen models see images? The Qwen3.8-27B and Qwen3.8-Flash-Next cards list the type as "Causal Language Model with Vision Encoder." The Qwen Cloud page lists image, text and video as inputs for Qwen3.8-Max.

Has Qwen 4 been released? Not as of November 4, 2026. Alibaba's latest written statement on Qwen 4 that we found, from September 22, 2026, says Qwen 4 "is currently in training."

Sources

ChinaQwenExplainersOpen Weights

Written by

axc
axc

Edited by the llms.blog desk.

More in Qwen

All Qwen →

Popular this week