What Is Qwen AI? Owner, Models, Access & Open Weights

Last verified: August 23, 2026

Independent and unofficial: Qwen AI Chat is not affiliated with, endorsed by or operated by Alibaba, Alibaba Cloud or the Qwen team. This article is an independent guide checked against official Qwen and Alibaba sources. Any chat demo on this website may use a different model or provider and should not be mistaken for the official Qwen Studio service.

Qwen AI is a family of language and multimodal artificial-intelligence models developed by Alibaba’s Qwen team. Qwen is not one model, one chatbot or one API. The name covers downloadable open-weight checkpoints, hosted models, visual and audio model families, coding tools and the official Qwen Studio assistant.

People can use Qwen through different products: the official Qwen Studio website and apps, Alibaba Cloud Model Studio or QwenCloud APIs, Qwen Code, or a specific checkpoint downloaded from the official Qwen organization on Hugging Face. Capabilities, context limits, prices, availability and licenses vary by model and access route.

The most important distinction: “Qwen” is the overall model ecosystem. “Qwen Studio” is an official hosted assistant. A Qwen API model is a hosted service with a specific model ID. An open-weight Qwen model is a particular downloadable checkpoint with its own repository and license. These terms are related, but they are not interchangeable.

Qwen AI at a glance

Question	Verified answer
What is Qwen AI?	A family of language and multimodal AI models and related products
Who develops Qwen?	Alibaba’s Qwen team; official sources describe Qwen as an Alibaba model family
Is Qwen one model?	No. Qwen includes multiple generations, sizes, architectures and specialist families
Is Qwen the same as Qwen Studio?	No. Qwen Studio is the official consumer assistant that provides access to hosted capabilities
Is every Qwen model open source?	No. Some checkpoints are released with downloadable weights; other models are hosted only or available as previews
Can Qwen run locally?	Some open-weight checkpoints can be self-hosted if the exact model’s license and hardware requirements are met
Does Qwen have an API?	Yes. Alibaba Cloud Model Studio and QwenCloud document hosted access, with model and endpoint availability varying by region
Does Qwen support images or audio?	Specific Qwen-VL, Qwen-Omni, Qwen-Audio and related models support additional modalities; do not apply that claim to every model
Does Qwen browse the web by itself?	No. Real-time data and external actions require product tools, retrieval or Function Calling implemented around the model

What does “Qwen” refer to?

Official Qwen materials use the English name Qwen and the Chinese name 通义千问 (Tongyi Qianwen). The official Qwen organization describes it as the large-language-model family built by Alibaba Cloud and says it releases language models, multimodal models and related projects. This article does not assign a literal English translation or an origin story to the name because the official technical sources reviewed here do not establish one clearly.

The term “Qwen AI” is commonly used by searchers as an umbrella phrase. Depending on the context, a person may be looking for the official chat assistant, a model such as Qwen3, an API alias such as Qwen-Plus, a vision model, or downloadable model weights. Accurate documentation should always identify which of those it means.

Qwen, Qwen Studio, APIs and open weights

Name	What it is	How it is accessed
Qwen	The overall Alibaba model family and ecosystem	Through several products, APIs and repositories
Qwen Studio	The official hosted consumer assistant	Web interface and official apps
Qwen API / Model Studio / QwenCloud	Hosted developer services exposing supported model IDs and tools	Region- and product-specific API keys and endpoints
Open-weight Qwen models	Specific checkpoints whose weights can be downloaded under their stated licenses	Official Hugging Face, ModelScope and GitHub pages
Qwen Code	An official open-source terminal coding agent	The official Qwen Code repository and supported model providers

A response produced in Qwen Studio does not prove that the same model ID, context window or tools are available through an API. Likewise, a downloadable checkpoint does not prove that it is the model behind the current Qwen Studio interface. Product documentation must be checked separately.

For the checkpoint-by-checkpoint distinction between Apache-licensed weights, custom or research licenses, and hosted-only service IDs, read Is Qwen AI open source?.

Who owns and develops Qwen AI?

Qwen is developed by Alibaba’s Qwen team. Official descriptions differ slightly by publication—some say Alibaba Cloud and newer project pages may say Alibaba Group—but they consistently identify Qwen as an Alibaba model family. Qwen AI Chat, the website publishing this guide, is an independent third-party site and does not own Qwen or represent Alibaba.

A corrected history of Qwen

Period	Verified milestone
August 2023	Alibaba released Qwen-7B and Qwen-7B-Chat. The official repository records a 2,048-token context for the earliest build.
September 2023	Qwen updated Qwen-7B from 2,048 to 8,192 tokens and published the Qwen Technical Report. The 2023 checkpoint should not be described as a 32K-at-launch model.
February–March 2024	Qwen1.5 introduced a wider set of model sizes and announced an MoE model. The dedicated Qwen1.5-MoE-A2.7B release followed in March, before Qwen2.
June 2024	Qwen2 launched with dense models and Qwen2-57B-A14B, an MoE model, plus broader multilingual and long-context support.
September 2024	Qwen2.5 launched with general, coding and mathematics model lines. Its official release date was September 19, 2024—not early 2025.
April 2025	Qwen3 launched with six dense open-weight models and two open-weight MoE models in the initial release.
2025–2026	The ecosystem expanded through later general families such as Qwen3-Next, plus coding, vision, omni-modal, image and hosted model lines. Release and access status must be checked for each exact name.
July 2026	`qwen3.8-max-preview` became available as a hosted preview through QwenCloud Token Plan and supported Qoder products. QwenCloud later retired the preview; the legacy ID now routes to production `qwen3.8-max`.
August 3, 2026	QwenCloud launched the production ID `qwen3.8-max` as its current flagship. The official release notes describe a native vision-language, 2.4-trillion-parameter Mixture-of-Experts model with hybrid thinking enabled by default and a 1M-token context window. These are provider-published specifications and positioning, not an independent benchmark result.
August 12, 2026	Qwen published `Qwen/Qwen3.8-2.4T-A95B` and an official FP8 variant. The repositories use the custom Qwen3.8-Max License and are distinct from the hosted API ID.
August 14, 2026	Qwen published the separate dense `Qwen/Qwen3.8-27B` checkpoint and FP8 variant under Apache 2.0. Its model card documents 262,144 native context and an optional 1M extension.
August 19, 2026	QwenCloud launched the hosted `qwen3.8-27b` service ID. The lowercase hosted ID and the `Qwen/Qwen3.8-27B` repository remain separate access routes.

Why Qwen-7B context figures can appear to conflict: the official project history distinguishes the original August build, the September update and later long-context extension work. The first build was 2K, the September checkpoint was updated to 8,192, and later Qwen materials described 32K capability for an extended Qwen-7B state. “The original Qwen-7B launched with 32K” is therefore inaccurate.

Did Mixture-of-Experts begin with Qwen2?

No. Mixture-of-Experts, usually abbreviated MoE, appeared in the Qwen family during Qwen1.5. The Qwen1.5 announcement included an MoE model, and Qwen later published Qwen1.5-MoE-A2.7B in March 2024. Qwen2, released in June 2024, also included an MoE checkpoint, Qwen2-57B-A14B, but it was not the beginning of MoE in Qwen.

In a dense model, the same main feed-forward parameters participate for each token. In an MoE model, a routing mechanism activates a subset of expert components. This can separate total parameters from active parameters, so labels such as “235B-A22B” should be read using the exact model card. The Qwen family contains both dense and MoE architectures; neither description applies to every Qwen model.

The main types of Qwen models

Model category	Typical role	Important qualification
General language models	Writing, summarization, extraction, translation, analysis and conversation	Reasoning modes, languages and context differ by exact model
Qwen-Coder	Code generation, explanation, editing and software-agent workflows	A coding model does not execute or modify systems unless a surrounding tool is allowed to do so
Qwen-VL	Understanding images, documents and, for supported models, video	Accepted formats, limits and visual capabilities are model- and provider-specific
Qwen-Omni and audio models	Supported combinations of text, image, audio, video and speech	Input and output modalities vary; “multimodal” does not guarantee every modality
Qwen-Image	Image generation and editing	This is a separate model line, not an automatic feature of every language model
Embedding and reranking models	Semantic search, retrieval and ranking	They normally return vectors or relevance scores rather than a chat answer
Hosted Max, Plus, Flash and other aliases	Managed API and product access	An alias may be updated or region-limited; it is not necessarily identical to a downloadable checkpoint

How do Qwen models work?

At a high level, Qwen language models are Transformer-based generative models trained to predict tokens and then post-trained for instruction following or other tasks. That broad explanation does not make every Qwen model architecturally identical. Model generations can differ in attention design, tokenizer, training mixture, context length, dense or MoE structure, multimodal encoders and post-training methods.

A Qwen model generates output from the context supplied to it. The application may add other components:

Retrieval-augmented generation (RAG): retrieves relevant documents or records and includes them as context.
Function Calling: lets the model request a declared function while the application validates and executes it.
Conversation state: the application stores or resends prior turns; the Qwen Chat Completions API is documented as stateless.
Safety and permissions: application rules, moderation, authentication and human review constrain what can be returned or acted upon.

These surrounding components should not be described as knowledge or powers built into the base model. Alibaba’s official Function Calling documentation explicitly says an LLM cannot access real-time data or external systems on its own.

Is Qwen open source?

The accurate answer is: some Qwen models have downloadable open weights, while other Qwen models are hosted services or previews. “Open source” is often used loosely in model marketing, but weights, inference code, training code and training data are separate components. For practical reuse, check the exact repository and license rather than assuming one license covers the entire family.

The initial Qwen3 open-weight release listed its dense and MoE models under Apache 2.0. Qwen3.8 then split across different licenses: the 2.4T-A95B repositories use the custom Qwen3.8-Max License, while Qwen/Qwen3.8-27B and its FP8 variant use Apache 2.0. License statements must therefore remain checkpoint-specific.
Other generations include model-specific exceptions. For example, Qwen’s Qwen2.5 announcement said its 3B and 72B variants did not use the same Apache 2.0 license as the other listed open models.
Hosted aliases such as Max, Plus or a preview ID are not automatically downloadable.
A planned open-weight release is not open weight until an official checkpoint and license are published.

Before commercial deployment, read the license file attached to the exact model repository and review any use restrictions that apply to your organization and jurisdiction.

How to use Qwen AI

Route	Best for	What to verify
Qwen Studio	Using the official hosted assistant without building an integration	Current selected model, product limits, terms and privacy policy
Alibaba Cloud Model Studio API	Applications needing managed API access	Region, Base URL, API key, exact model ID, pricing and supported features
QwenCloud plans	Supported interactive coding and agent tools	Plan-specific model allowlist and use policy; do not assume every plan allows application backends
Open-weight checkpoint	Self-hosting, research, fine-tuning or controlled deployment	License, model card, memory, precision, serving stack and security responsibilities
Qwen Code	Terminal-based coding assistance and agent workflows	Provider, authentication method, enabled tools and model configuration

Never invent an endpoint such as api.qwen.ai. Alibaba Cloud publishes different endpoints and API keys for different regions and products. Copy the current Base URL from the official guide for the account and region you actually use.

Current Qwen3.8 status: Max, 27B and the retired preview alias

As rechecked on August 23, 2026, the production model ID qwen3.8-max remains QwenCloud’s hosted flagship. It launched on August 3. QwenCloud describes it as a native vision-language Mixture-of-Experts model with 2.4 trillion total parameters, hybrid thinking enabled by default and a 1M-token context window. Those specifications and positioning statements come from the provider; this independent guide has not reproduced a universal performance ranking.

qwen3.8-max-preview has ended its preview period and is officially retired. QwenCloud temporarily continues to accept the old ID but routes those requests to qwen3.8-max, with Credits deduction and usage statistics recorded under the production model. New configurations should use qwen3.8-max directly.

Qwen publishes the related first-party Qwen/Qwen3.8-2.4T-A95B and Qwen/Qwen3.8-2.4T-A95B-FP8 checkpoints under the custom Qwen3.8-Max License. The hosted Max service and checkpoint IDs remain distinct: the model card says the hosted service adds capabilities such as vision input, non-thinking mode, a 1M default context and built-in tools.

Qwen also released the separate dense Qwen/Qwen3.8-27B and Qwen/Qwen3.8-27B-FP8 checkpoints on August 14 under Apache 2.0. The model card documents 262,144 native context and an optional 1M extension. On August 19, QwenCloud launched the hosted qwen3.8-27b ID; that lowercase service name is not the same artifact as the Hugging Face repository.

See the independent Qwen3.8 Max guide, API guide, and Qwen download directory for route-specific details.

What can Qwen AI be used for?

Drafting, rewriting, summarizing and extracting text.
Multilingual assistance with quality testing for each required language.
Code explanation and generation with suitable Qwen-Coder or general models.
Document and image analysis with a supported visual model.
Speech or audio workflows with an exact model that documents those modalities.
Search and question answering when connected to a controlled retrieval system.
Agents that request tools through Function Calling, with execution controlled by the application.
Embedding and reranking for semantic retrieval and recommendation pipelines.

These are possible application categories, not guarantees of accuracy, productivity, revenue or suitability. Evaluate the exact model on representative data before deployment.

What Qwen cannot safely guarantee

Factual accuracy: generated answers can be wrong, unsupported or outdated.
Current information: real-time prices, news, inventory or private records require an external source.
Perfect retrieval: RAG can improve grounding but can retrieve the wrong passage or miss the right one.
Correct tool use: function names and arguments must be validated before execution.
Professional judgment: medical, legal, financial and safety-critical outputs require qualified review.
Identical provider behavior: an API, Qwen Studio and a self-hosted checkpoint can have different settings and tools.
Permanent aliases: hosted “latest” and preview models may be updated, deprecated or replaced.

How to choose a Qwen model

Define the task and input type. Decide whether the application needs text, images, audio, code, embeddings or tool use.
Choose hosted or self-hosted access. Compare operational control, engineering effort, data handling and cost.
Verify the exact model card. Check context, output limits, supported features, license and recommended serving software.
Confirm region and endpoint. Alibaba Cloud regions can differ in model IDs, Base URLs, keys, features and pricing.
Build a representative evaluation. Test accuracy, unsupported claims, latency, cost, safety and failure recovery.
Version the deployment. Record the model ID, snapshot where available, prompt, tools and evaluation date.

Accuracy, privacy and security

Data handling depends on the product used. Alibaba Cloud Model Studio states that transmitted application and training data is encrypted and is not used for model training, but that statement belongs to Model Studio and should not automatically be applied to Qwen Studio, a third-party provider or an independent website. Review the applicable product terms, region and privacy documentation before sending personal, confidential or regulated information.

Keep API keys on the server and out of browser code, prompts, screenshots and repositories.
Send only the data required for the task and remove unnecessary personal information.
Authenticate users before exposing private documents or account-specific tools.
Treat model output and tool arguments as untrusted input.
Use allowlists, schemas, permission checks and explicit confirmation for state-changing actions.
Log model and tool versions so incidents can be reproduced and investigated.

Frequently asked questions

What is Qwen AI in simple terms?

Qwen AI is Alibaba’s family of generative language and multimodal models. It includes general models, specialist models, hosted services, official applications and downloadable checkpoints.

Who owns Qwen AI?

Qwen is an Alibaba model family developed by the Qwen team. This independent website is not Alibaba and does not own or operate the official Qwen products.

Is Qwen AI the same as Qwen Studio?

No. Qwen is the overall model family. Qwen Studio is the official hosted assistant through which users can access selected Qwen capabilities.

Is Qwen AI free?

There is no single family-wide answer. Qwen Studio may offer access subject to its current terms and limits. Downloadable weights may be available without a model fee but require suitable computing resources and compliance with the model license. Hosted APIs have provider-, model- and region-specific pricing.

Is every Qwen model open source?

No. Some Qwen checkpoints have downloadable weights under a stated license, while other models are hosted only or remain previews. Check the exact official repository and license.

Can I run Qwen locally?

You can self-host a Qwen checkpoint when official weights are available, the license permits your use, and your hardware and inference stack support it. Hosted-only models and unreleased previews cannot be downloaded from an official checkpoint.

Does Qwen remember conversations?

The Qwen Chat Completions API is documented as stateless. An application must maintain and resend the necessary conversation history or use a supported state-management product. Qwen Studio may manage session history as a product feature.

When did Qwen first use an MoE architecture?

Qwen1.5 included an MoE model, and Qwen1.5-MoE-A2.7B was published in March 2024. This predates the June 2024 Qwen2 release.

Did the original Qwen-7B launch with 32K context?

No. Qwen’s project history records an initial 2,048-token build in August 2023 and an update to 8,192 tokens in September. The 32K figure appears in later extended Qwen-7B materials and should not be assigned to the original launch.

What is the difference between qwen3.8-max and qwen3.8-max-preview?

qwen3.8-max is the production flagship ID. qwen3.8-max-preview is retired, but QwenCloud temporarily accepts the legacy ID and routes it to qwen3.8-max; Credits and usage statistics are recorded under the production model. Update integrations to use qwen3.8-max directly, and do not infer Qwen Studio’s selected model from API documentation.

Official sources

Explore Qwen models and developer access

Browse Qwen Models

Read the Qwen API Guide

View Download Options

Editorial policy: This page distinguishes original releases from later updates, hosted products from downloadable weights, and vendor statements from independently established facts. Recheck model status, licenses, endpoints and product defaults before each material update.