Last verified: July 26, 2026.
Correct answer: Do not write that “all Qwen2.5 and Qwen3 family models are Apache-2.0.” The official Qwen2.5 table assigns the Qwen Research License to 3B and the custom Qwen License to 72B. The initial downloadable Qwen3 checkpoints use Apache 2.0, but that does not transfer the license to every hosted Qwen3-branded Max, Plus, or Flash service.[1][15]
Qwen-AI.chat is an independent informational website. It is not operated, authorized, or endorsed by Alibaba Cloud or the Qwen Team. This page explains public Qwen model and service documentation; it does not grant a license or replace review of the applicable legal terms.
Qwen-AI.chat is an independent informational website. It is not operated, authorized, or endorsed by Alibaba Cloud or the Qwen Team. This page explains public Qwen model and service documentation; it does not grant a license or replace review of the applicable legal terms.
Qwen open-source status at a glance
| Qwen component | Accurate classification | What to verify |
|---|---|---|
| Checkpoint with public weights and Apache 2.0 | Apache-2.0-licensed open weights | The exact official repository and its LICENSE file |
| Checkpoint under Qwen Research License | Downloadable, research-only weights | Non-commercial scope and separate-license requirement |
| Checkpoint under Qwen License | Public weights under a custom license | Redistribution, attribution, scale, and other conditions |
| Max, Plus, Flash, or another Model Studio alias | Hosted model service unless an exact weight repository is separately published | Provider terms, model ID, region, pricing, and lifecycle |
| Qwen Chat | Hosted application | Application terms and privacy information—not a model-weight license |
The practical answer to “Is Qwen AI open source?” is therefore model-specific. A public download is not proof of Apache 2.0, and access through an API is not proof that the underlying weights are public.
Open source, open weights, and hosted access are different
Qwen’s own announcements have used both “open source” and “open weight.” For licensing accuracy, open weight is the safer umbrella term when an official repository publishes trained model parameters. The Open Source Initiative distinguishes open weights from a complete Open Source AI system, which also involves access to information and code needed to understand, modify, and study the system.[20]
| Term | Meaning in this guide |
|---|---|
| Open weights | The trained parameters can be downloaded from an official repository. |
| Apache-2.0-licensed open weights | The exact checkpoint’s files identify Apache License 2.0. |
| Research-only weights | The weights are downloadable, but commercial use is not granted by the published research license. |
| Custom-licensed open weights | The weights are downloadable under Qwen’s own license rather than Apache 2.0. |
| Hosted model | A provider runs the model and exposes an API or interface; users do not necessarily receive the weights. |
Model implementation code can also have a different scope from model weights. An Apache-licensed GitHub repository does not automatically assign Apache 2.0 to every checkpoint, hosted alias, training dataset, or product using that code.
Official Qwen2.5 license table
The official Qwen2.5 LLM table lists seven model sizes and does not use one license across the family.[1] The same size-level split applies to the main base and instruction-tuned releases, but the exact repository remains the final reference.
| Core Qwen2.5 size | Official license | Publish-safe classification |
|---|---|---|
| 0.5B | Apache 2.0 | Apache-2.0-licensed open weights |
| 1.5B | Apache 2.0 | Apache-2.0-licensed open weights |
| 3B | Qwen Research License | Research-only downloadable weights |
| 7B | Apache 2.0 | Apache-2.0-licensed open weights |
| 14B | Apache 2.0 | Apache-2.0-licensed open weights |
| 32B | Apache 2.0 | Apache-2.0-licensed open weights |
| 72B | Qwen License | Downloadable weights under a custom license |
The Qwen/Qwen2.5-3B license limits the granted rights to non-commercial research or evaluation and says commercial users must request a separate license.[2] The 72B repository is publicly downloadable, but its custom Qwen License—not Apache 2.0—governs use.[3]
Important correction: Qwen2.5-72B is not a hosted-only model. Its weights are public, but they use the Qwen License. “Downloadable” and “Apache-licensed” are separate questions.
Licenses for Qwen2.5 Coder, Math, VL, Omni, and QwQ
The Qwen2.5 name also covers specialist families. Their licenses follow the exact checkpoint, not a marketing-family label.
| Family or checkpoint | Official license | Accurate treatment |
|---|---|---|
| Qwen2.5-Coder 0.5B, 1.5B, 7B, 14B, and 32B | Apache 2.0 | Apache-2.0-licensed open weights |
| Qwen2.5-Coder 3B | Qwen Research License | Research-only downloadable weights |
| Qwen2.5-Math 1.5B and 7B | Apache 2.0 | Apache-2.0-licensed open weights |
| Qwen2.5-Math 72B | Qwen License | Custom-licensed open weights |
Qwen2.5-VL-3B-Instruct | Qwen Research License | Research-only downloadable weights |
Qwen2.5-VL-7B-Instruct and Qwen2.5-VL-32B-Instruct | Apache 2.0 | Apache-2.0-licensed open weights |
Qwen2.5-VL-72B-Instruct | Qwen License | Custom-licensed open weights—not a hosted-only “VL-Max” checkpoint |
Qwen2.5-Omni-3B | Qwen Research License | Research-only downloadable weights |
Qwen2.5-Omni-7B | Apache 2.0 | Apache-2.0-licensed open weights |
Qwen/QwQ-32B | Apache 2.0 | Apache-2.0-licensed open reasoning-model weights |
This corrects two common errors. First, Qwen2.5-Coder-3B is not Apache 2.0 even though the 32B Coder checkpoint is. Second, Qwen2.5-VL-72B-Instruct has public weights under the Qwen License; it should not be conflated with a separately named hosted VL-Max service.
For modality and deployment details, see our Qwen2.5-Omni guide. For the reasoning checkpoint, see QwQ-32B. The broader Qwen2.5 family guide owns general architecture and capability coverage; this page owns licensing and openness.
Which Qwen3 models are Apache 2.0?
Qwen’s April 2025 release made eight Qwen3 checkpoints available as open weights under Apache 2.0: six dense models and two mixture-of-experts models.[15] The official Qwen3 repository states that its open-weight models use Apache 2.0.[16]
| Qwen3 release group | Published checkpoints | License |
|---|---|---|
| Dense Qwen3 | 0.6B, 1.7B, 4B, 8B, 14B, and 32B | Apache 2.0 |
| MoE Qwen3 | 30B-A3B and 235B-A22B | Apache 2.0 |
Later Qwen3.x releases also include public Apache-2.0 checkpoints. Verified examples include Qwen/Qwen3.5-397B-A17B and Qwen/Qwen3.6-35B-A3B.[17][18] However, the safe statement is still “Qwen’s officially published open-weight checkpoints use the license shown in their repository,” not “every service whose name starts with Qwen3 is Apache 2.0.”
See the Qwen3 model guide for architecture, thinking modes, and model selection. License verification should remain checkpoint-specific.
Why Max, Plus, and Flash should be treated separately
Alibaba Cloud Model Studio documents names such as qwen3.7-max, qwen3.7-plus, qwen3.7-flash, qwen-max, qwen-plus, and qwen-flash as hosted model IDs.[19] An API alias is not a Hugging Face checkpoint and does not, by itself, provide a model-weight license.
The same distinction applies to Qwen2.5-Max. Its official January 2025 announcement directs users to Qwen Chat and the Alibaba Cloud API but does not announce a downloadable Qwen2.5-Max checkpoint.[22] That is evidence for hosted access, not an Apache 2.0 weight release.
| Question | Downloadable checkpoint | Hosted service |
|---|---|---|
| Where is it accessed? | Official model repository | Provider API or web application |
| What governs the weights? | The repository’s model LICENSE | No weight license is implied if weights are not distributed |
| What governs API use? | Not applicable to local inference | Provider service terms, pricing, quotas, regions, and data policies |
| Can it be self-hosted? | Yes, subject to license and hardware | Not from the API alias alone |
A public checkpoint can also be offered through Model Studio. In that case, the two routes remain legally distinct: downloading the weights follows the repository license, while calling the managed endpoint follows the platform’s service terms. See the Qwen API guide for hosted integration and the Qwen API pricing guide for provider costs.
What each Qwen license means for commercial use
| License or access route | Commercial-use summary | Key caution |
|---|---|---|
| Apache License 2.0 | Generally permits commercial use, modification, and distribution subject to its conditions | Keep required license and notices, mark modified files, respect patent terms, and do not imply trademark rights |
| Qwen Research License | The published grant is limited to non-commercial research or evaluation | Commercial use requires a separate license from Alibaba Cloud |
| Qwen License | Grants use, modification, and distribution rights subject to custom conditions | A commercial product or service exceeding 100 million monthly active users must request a license; redistribution and attribution requirements also apply |
| Model Studio hosted access | Commercial API use depends on the provider plan and terms | Do not substitute a checkpoint license for API terms or vice versa |
Apache 2.0 is permissive, but “permissive” does not mean “without conditions.” It also does not grant ownership of the Qwen trademark, guarantee non-infringement, remove regulatory duties, or make model output automatically safe to use.
How to verify any Qwen model before deployment
- Record the exact identifier. For example,
Qwen/Qwen2.5-3B-InstructandQwen/Qwen2.5-7B-Instructdo not share the same license. - Use an official Qwen repository. A community quantization or fine-tune can introduce additional files or terms.
- Open the actual
LICENSEfile. Do not rely only on a search snippet, family name, blog summary, or neighboring model. - Check the model card and repository revision. License metadata or files can be corrected, so record the verification date.
- Trace derivatives to the parent model. A GGUF, AWQ, adapter, or fine-tune still inherits obligations from the underlying checkpoint.
- Separate weights from services. If the identifier is a Model Studio API alias, review provider terms rather than assuming an Apache weight release exists.
- Review the intended use. Commercial distribution, training another model, high-scale deployment, regulated data, and trademark use can trigger different obligations.
Recommended editorial rule: Never assign a license from the words “Qwen2.5,” “Qwen3,” “Max,” “Plus,” or “Flash” alone. Publish the exact checkpoint or service ID, the exact license or access terms, a primary-source link, and a “Last verified” date.
Official sources
- Official Qwen2.5 LLM model and license table
- Qwen2.5-3B Qwen Research License
- Qwen2.5-72B-Instruct Qwen License
- Official Qwen2.5-Coder license table
- Qwen2.5-VL-3B-Instruct license
- Qwen2.5-VL-7B-Instruct official repository
- Qwen2.5-VL-32B-Instruct official repository
- Qwen2.5-VL-72B-Instruct Qwen License
- Qwen2.5-Omni-3B license
- Qwen2.5-Omni-7B Apache 2.0 license
- Qwen2.5-Math-1.5B-Instruct official repository
- Qwen2.5-Math-7B-Instruct official repository
- Qwen2.5-Math-72B-Instruct official repository
- Official QwQ-32B release and license statement
- Official Qwen3 open-weight release
- Official Qwen3 repository and license statement
- Qwen3.5-397B-A17B official repository
- Qwen3.6-35B-A3B official repository
- Alibaba Cloud Model Studio model catalog
- Open Source AI Definition 1.0
- Apache License 2.0
- Official Qwen2.5-Max announcement
This page was last cross-checked against the official sources above on July 26, 2026.
Frequently asked questions
Is Qwen AI open source?
Partly. Qwen includes many publicly downloadable checkpoints, several under Apache 2.0, but it also includes research-only weights, custom-licensed weights, hosted APIs, and the Qwen Chat application. Check the exact checkpoint or service rather than applying one license to the whole ecosystem.
Are all Qwen2.5 models licensed under Apache 2.0?
No. The official core Qwen2.5 table lists 0.5B, 1.5B, 7B, 14B, and 32B under Apache 2.0; 3B uses the Qwen Research License and 72B uses the custom Qwen License. Specialist families also have model-specific exceptions.
Can Qwen2.5-3B be used commercially?
The published Qwen Research License grants non-commercial research or evaluation use and says commercial users must request a separate license. Review the complete license before deployment.
Is Qwen2.5-72B downloadable?
Yes. Qwen2.5-72B weights are publicly downloadable, but they are governed by the custom Qwen License rather than Apache 2.0. Public weights do not automatically mean an Apache license.
Is Qwen2.5-VL-72B a hosted-only model?
No. Qwen2.5-VL-72B-Instruct has public weights under the Qwen License. It should not be conflated with a separately named hosted VL-Max service.
Are all Qwen3 models Apache 2.0?
The officially published open-weight Qwen3 checkpoints covered here use Apache 2.0, including the initial dense and MoE release. However, that does not prove that every Qwen3-branded hosted Max, Plus, or Flash service has downloadable Apache-licensed weights.
Are Qwen Max, Plus, and Flash open source?
These names commonly identify hosted Model Studio services. Do not call a service open-weight or Apache-licensed unless Alibaba also publishes an exact downloadable checkpoint with an explicit license file.
Can Apache-licensed Qwen models be used commercially?
Apache 2.0 generally permits commercial use, modification, and distribution subject to its conditions, including license and notice obligations. It does not grant trademark rights or remove other legal and regulatory duties.
What is the difference between Qwen open source and open weight?
Open weight means the trained parameters are downloadable. A complete Open Source AI system can require broader access to code and information about how the system was built. This guide uses open weight when discussing downloadable Qwen checkpoints.
How do I verify a Qwen model’s license?
Copy the exact model ID, open its official Qwen repository, inspect the LICENSE file and model card, trace any quantization or fine-tune to its parent checkpoint, and record the date reviewed. For API aliases, review the provider’s service terms instead.