Download Qwen: Official Apps and Model Weights

Last verified: August 3, 2026. Platform availability, model collections, repository ownership, and license notices can change. Recheck the linked first-party page before installing software or deploying a checkpoint.

To download Qwen safely, first decide whether you need the official Qwen Studio application or downloadable model weights. The app is a client for Qwen’s hosted service. Model weights are checkpoint files used for local or self-hosted inference. API access is a third option and does not provide model files.

Independence notice: Qwen-AI.chat is an independent informational website and demo. It is not operated, authorized, or endorsed by Alibaba, Alibaba Cloud, or the Qwen team. We do not host or redistribute the installers and weights linked on this page, and this directory does not imply that our chat interface runs any listed checkpoint.

Choose the correct Qwen download

Your goal	Correct destination	What you receive
Use Qwen in a browser	Qwen Studio Web	A hosted chat service; no model weights
Install the Qwen app	Official Qwen download page	The platform-specific app or store route
Run a model locally or on your own server	Qwen on Hugging Face or Qwen on ModelScope	First-party checkpoints, model cards, and license files
Find reference code and tools	QwenLM on GitHub	Official source repositories; weights may be linked separately
Call a hosted model from an application	Qwen API guide	Hosted API access; no downloadable checkpoint is implied

This is a curated directory of official distribution hubs, not a claim that every Qwen product has downloadable weights. A model shown in Qwen Studio or an API catalog is downloadable only when Qwen publishes a first-party weight repository for that exact model ID.

If you are deciding between hosted API access and self-hosting, use the Qwen pricing guide to compare provider token charges with GPU, storage, bandwidth, engineering, monitoring, security, and idle-capacity costs.

Download the official Qwen app

The official Qwen download destination lists Qwen Studio for iOS, Android, macOS, Windows, and the web. Use that central page instead of searching an app store by name: store titles, publishers, regional availability, and compatible devices can change, while search results may include unrelated or unofficial products.

• Web: opens the hosted Qwen Studio service in a browser.
• iOS and Android: follow the store route supplied by Qwen for your device and region.
• macOS and Windows: use the desktop option supplied on the official download page.
• Linux: the verified platform list on the central Qwen Studio download page does not include a Linux desktop app. Browser access and local open-weight deployment are separate alternatives.

Before installing, confirm that you began on the qwen.ai domain. If the page sends you to an app store, verify that the listing is the destination opened by that official page. Avoid third-party APK sites, repackaged desktop clients, and ads that imitate a download button.

Download Qwen Code: an agent, not model weights

Qwen Code is the official open-source terminal coding agent. Installing it gives you an application that can connect a model to files and tools; it does not download a Qwen model checkpoint automatically. Follow the installation and authentication instructions in the official repository, review requested permissions, and confirm the configured model endpoint before granting repository or command access.

Test boundary: we verified the official repository destination and product distinction, but we did not perform a clean-machine installation or publish a local performance result in this review. See Qwen solutions by use case for a safer coding-agent evaluation checklist.

Download official Qwen model weights

For local inference, fine-tuning, evaluation, or self-hosting, start from Qwen’s organization page on Hugging Face or ModelScope. Both hubs contain first-party model repositories under the Qwen owner name. Qwen’s documentation provides model-specific installation and inference instructions, including downloads from both hubs.

Qwen3.8 Max is hosted access, not a weight download

QwenCloud launched the production model ID qwen3.8-max on August 3, 2026. The release discloses a 2.4-trillion-parameter Mixture-of-Experts architecture and a 1M-token context window, but an architecture disclosure is not a download. The first-party release notes and model directory reviewed for this update do not publish a weight repository or license for that exact ID.

The production ID qwen3.8-max and the Token Plan preview ID qwen3.8-max-preview are hosted service identifiers. Neither should be converted into a Hugging Face path, local-runtime command, or hardware requirement unless Qwen publishes an exact first-party checkpoint and license. For hosted use, consult the Qwen3.8 Max guide and Qwen API guide.

Newest general first-party weights: Qwen3.6

At this page’s August 2, 2026 source check, Qwen3.6 is the newest general first-party open-weight Qwen generation with public standard checkpoints. Use the exact Qwen-owned repository ID; do not add an -Instruct suffix that the official Qwen3.6 IDs do not contain.

Exact first-party ID	Published form
Qwen/Qwen3.6-35B-A3B	Standard mixture-of-experts checkpoint: 35B parameters in total, 3B activated
Qwen/Qwen3.6-35B-A3B-FP8	First-party FP8 version of the 35B-A3B checkpoint
Qwen/Qwen3.6-27B	Standard dense 27B checkpoint
Qwen/Qwen3.6-27B-FP8	First-party FP8 version of the dense 27B checkpoint

The standard model cards describe post-trained multimodal checkpoints with a vision encoder, Apache-2.0 licensing, a native 262,144-token context, and a documented extension procedure up to 1,010,000 tokens. These specifications do not prove that a checkpoint fits your hardware or performs reliably on your workload.

Hosted IDs such as qwen3.8-max, qwen3.8-max-preview, and qwen3.7-plus are service targets, not public checkpoint names. At this page’s August 3, 2026 source check, Qwen3.6 remained the newest general first-party open-weight generation with standard checkpoints in Qwen’s official model organization. Recheck the Qwen models directory and the exact Qwen-owned repository before changing that statement.

The official Qwen3.6 repository documents current routes for Transformers, llama.cpp, MLX, SGLang, and vLLM. Verification boundary: we confirmed the first-party destinations and model-card claims, but we did not download the weights, install a runtime, publish a benchmark, or validate local hardware in this review.

API-only model names are not download links. A managed provider model such as qwen3.7-plus may be available for inference without a corresponding public checkpoint under that identifier. Only claim a model is downloadable when an official Qwen repository publishes the weights and license for that exact release.

Do not select a checkpoint by parameter count alone. The exact repository name can distinguish a base model from an instruction-tuned, thinking, multimodal, coding, embedding, reranking, quantized, or reduced-precision variant. The model card should state the task, expected software versions, input and output types, and any special serving requirements.

First-party weight checklist

1. Confirm that the repository owner is Qwen on Hugging Face or ModelScope.
2. Copy the complete checkpoint ID, including every suffix such as Instruct, Thinking, Base, AWQ, GGUF, or FP8.
3. Read the model card and the repository’s LICENSE file before downloading or redistributing the files.
4. Follow the dependency and serving instructions for that exact checkpoint rather than instructions written for another Qwen generation.
5. Record the repository revision and file manifest if you need a reproducible production deployment.

A format label is not proof that a file is official. Qwen publishes some first-party quantized and reduced-precision variants, while independent users publish others. If the owner is not Qwen, treat the repository as a community conversion and verify its source model, conversion method, license inheritance, and integrity separately.

For the supported download and inference workflow, consult the official Qwen Transformers documentation. Hardware requirements depend on the checkpoint, precision, context length, runtime, and workload; a single universal VRAM figure would be misleading.

Official Qwen weight collections

Qwen’s Hugging Face organization maintains collection pages that group related first-party repositories. Use the organization’s collection index when you need to see newly published families or specialist models without relying on a static third-party list.

Collection	What the official collection contains	Use it when
Qwen3.6	First-party Qwen3.6 checkpoints and published variants	You need a Qwen3.6 model card or weight repository
Qwen3.5	First-party Qwen3.5 checkpoints and published variants	You need a Qwen3.5 model card or weight repository
Qwen3	Dense and mixture-of-experts checkpoints, plus official variants	You need the original Qwen3 generation or its dated updates
Qwen2.5	Base and instruction-tuned language models across documented sizes	You need the general Qwen2.5 language-model family
QwQ	QwQ-32B, QwQ-32B-Preview, and first-party published variants	You specifically need the QwQ reasoning-model line

Specialist families have their own official collections. For example, use the dedicated pages for Qwen2.5-Omni, Qwen2.5-VL, Qwen2.5-Coder, or Qwen3-Coder. Our Qwen model directory explains model roles before you choose a checkpoint.

Historical Qwen model archive

“Archive” here means a stable first-party collection of earlier releases; it does not automatically mean that every checkpoint is deprecated or unsupported. Historical models remain useful for reproducibility, comparisons, compatibility work, and projects already built around their tokenizer or architecture.

Family	Official archive	Accuracy note
Original Qwen	First-generation Qwen collection	Includes Qwen-1.8B, 7B, 14B, and 72B lines with base, chat, or published quantized repositories.
Qwen1.5	Qwen1.5 collection	A separate generation; do not mix its model IDs or loading instructions with original Qwen.
Qwen2	Qwen2 collection	Contains base, instruct, MoE, quantized, and related Qwen2 repositories.
Qwen2.5	Use the general and specialist collections listed above.	Licenses and modalities must be checked on the exact checkpoint card.
Qwen3	Use the Qwen3 collection listed above.	Base, instruct, thinking, MoE, and format variants are not interchangeable.

Correction: Qwen-1.8B has 1.8 billion parameters

Qwen-1.8B means 1.8 billion parameters, not 18 billion. The official model card describes it as the “1.8B-parameter version” of the original Qwen family. Its official repository ID is Qwen/Qwen-1_8B, using an underscore in the model ID. See the official Qwen-1.8B model card before using this historical checkpoint.

The original Qwen-1.8B card links to the Tongyi Qianwen Research License Agreement. Its grant covers non-commercial research or evaluation, while commercial use requires a separate license. Those terms are specific to this historical repository and must not be generalized to other Qwen checkpoints.

Licenses are model-specific

Do not assume that one license covers an entire Qwen generation, every parameter size, every modality, or every quantized derivative. The authoritative terms are the license file and model card in the exact repository you download. Our What Is Qwen AI guide explains the distinction between open weights, open-source code, hosted services, and checkpoint-specific licenses.

Artifact	Terms to read
Qwen Studio app or web service	The service terms and privacy policy linked by Qwen Studio
Model checkpoint	The model card and LICENSE in that exact weight repository
Source code or tool	The license in the specific QwenLM GitHub repository
Alibaba Cloud Model Studio API	The hosted-service documentation, account terms, and applicable model terms

Names such as Max, Plus, or Flash may identify hosted products. Their presence in a chat or API catalog does not prove that downloadable weights exist. Require a first-party model repository for the exact name before describing it as downloadable.

Before redistributing or deploying any checkpoint, verify its exact repository license in the Qwen open-source and open-weight guide; a hosted model ID is not a download license.

A safe Qwen download process

1. Choose the artifact type: hosted app, model weights, source code, or API.
2. Start from a first-party hub: Qwen’s download page, Qwen Studio, the Qwen organizations on Hugging Face or ModelScope, or QwenLM on GitHub.
3. Verify the exact identity: owner, complete checkpoint name, modality, tuning type, precision, and repository revision.
4. Read the terms: check the license and usage documentation that belongs to that artifact.
5. Follow the model card: use its dependency, memory, and runtime guidance instead of a generic hardware promise.
6. Test before deployment: evaluate quality, safety, latency, and resource use with your own workload.

Downloading weights does not automatically create a working chat application. You still need a compatible inference runtime, sufficient memory and storage, a chat template when required, and appropriate application-level safeguards. For hosted access instead, use the API documentation guide linked in the selection table.

Frequently asked questions

Official sources reviewed

• QwenCloud production qwen3.8-max release
• QwenCloud text-generation model directory
• Official Qwen model organization on Hugging Face
• Qwen official download destination — app and web platform routes; linked in the app section.
• Qwen Studio — official hosted chat service; linked in the selection table.
• Qwen organizations on Hugging Face and ModelScope — first-party models and collections; linked in the selection table.
• QwenLM on GitHub — official source repositories and project links; linked in the selection table.
• Qwen documentation — official inference and deployment guidance; linked in the weights section.
• Qwen3.6 official repositories and model cards — exact standard and FP8 IDs, architecture, context, license, and current runtime guidance; linked in the Qwen3.6 section.
• Qwen-1.8B model card and research license — parameter count, repository identity, and historical terms; linked in the correction section.

---

Editorial review: The Qwen-AI.chat Editorial Team matched the platform list, organization ownership, model IDs, collection scope, and license warnings to the first-party sources above. Corrections can be submitted through our contact page.