Qwen is Alibaba's answer to a simple question: what happens when one of the world's largest e-commerce and cloud companies gives away its models? In three and a half years, the Tongyi Qianwen chatbot of 2023 grew into Qwen, the most widely fine-tuned open model family in the world, and then into Qwen 3.7 Max and Qwen 3.8, models built for long agent tasks.
How did Alibaba decide which Qwen models to open and which to sell? Why did developers pick Qwen as their default base model? And what changed between Qwen 2.5, Qwen 3, and Qwen 3.8? This is the full Qwen story, updated for October 2026.
TL;DR: Alibaba launched Tongyi Qianwen in April 2023 and released open Qwen weights from August 2023. Qwen3 (April 2025) shipped under Apache 2.0 in eight sizes, Qwen 3.7 Max (May 2026) targeted agent work, and Qwen 3.8 (August 2026) brought a 2.4-trillion-parameter open model. Qwen 3.7 Plus is in the Taskade model picker on every paid plan. Build an app with it →
The TSK-1 Benchmark sends the same app request to each model family and records what got built. The Qwen results show how Qwen models do on real app builds.
What Is Qwen?
Qwen is the family of large language models built by Alibaba Cloud, the cloud computing arm of Alibaba Group in Hangzhou, China. The name is short for Tongyi Qianwen. Alibaba ships some Qwen models as open weights, most under Apache 2.0, and keeps its largest Max, Plus, and Flash models as paid services on Alibaba Cloud.
That split is the core of the Qwen strategy. Open models win developers, research citations, and goodwill. Paid flagships earn cloud revenue. Each new generation feeds both sides, because the same research produces a free small model and a paid large one.
Qwen at a Glance (2026)
| Fact | Detail |
|---|---|
| Maker | Alibaba Cloud (Alibaba Group) |
| Chinese name | Tongyi Qianwen |
| First beta | April 2023 |
| First open weights | Qwen-7B, August 2023 |
| Signature open release | Qwen3, April 2025, Apache 2.0, 0.6B to 235B |
| Newest flagships | Qwen 3.7 Max (May 2026), Qwen 3.8 (August 2026) |
| Leadership | Alibaba Token Hub group, led by CEO Eddie Wu, since March 2026 (after Lin Junyang) |
| In Taskade | Qwen 3.7 Plus on every paid plan, Qwen 3.7 Max on Max and Enterprise (details) |
What Does Tongyi Qianwen Mean?
Tongyi is the brand Alibaba uses for its AI products, and Qianwen means "a thousand questions." Together the name describes a model that can answer whatever you ask it. Qwen is the short English name, formed from the first letters of Qianwen, and it is the name the open-source community uses.
Why Did Alibaba Build Qwen?
Alibaba built Qwen because generative AI threatened to change both of its core businesses. Alibaba Cloud is the largest cloud provider in China, and every company that wanted AI would need models and the compute to run them. Alibaba's shopping apps needed AI search, product copy, and customer service at a scale few companies face.
When ChatGPT launched in late 2022, Chinese tech companies raced to answer it. Alibaba announced Tongyi Qianwen in April 2023 and opened a beta to business users. By September 2023, after Chinese regulators approved public generative AI services, it opened the chatbot to everyone.
Every Qwen Generation Explained (2023 to 2026)
Alibaba has released several hundred Qwen checkpoints across text, code, math, vision, audio, and reasoning. The table lists the generations that changed the family's direction.
| Date | Release | Why it mattered |
|---|---|---|
| Apr 2023 | Tongyi Qianwen beta | Alibaba's first chatbot, business users only |
| Aug 2023 | Qwen-7B open weights | First open Qwen model |
| Sep 2023 | Public launch, Qwen-14B | Open to everyone after regulatory approval |
| Dec 2023 | Qwen-72B | First large open Qwen model |
| Jun 2024 | Qwen2 | Most sizes moved to Apache 2.0 |
| Sep 2024 | Qwen2.5 | Seven sizes from 0.5B to 72B, plus Coder and Math |
| Nov 2024 | QwQ-32B-Preview | First open Qwen reasoning model |
| Jan 2025 | Qwen2.5-Max | Large proprietary MoE flagship |
| Apr 2025 | Qwen3 | Eight open sizes, hybrid thinking, Apache 2.0 |
| Jul 2025 | Qwen3-Coder | 480B open coding model for agents |
| Sep 2025 | Qwen3-Max | Over 1 trillion parameters, API-only |
| Feb 2026 | Qwen3.5 | Open 397B MoE, plus the proprietary 3.5-Plus |
| Apr 2026 | Qwen3.6 | Open 35B MoE under Apache 2.0, plus 3.6-Plus |
| May 2026 | Qwen 3.7 Max | Agent flagship with a 1-million-token context |
| Jun 2026 | Qwen 3.7 Plus | Lower-cost multimodal sibling |
| Aug 2026 | Qwen 3.8 | 2.4T open flagship weights, plus Qwen3.8-27B |
Qwen-7B to Qwen-72B (2023): Opening the Weights
Alibaba released Qwen-7B as open weights in August 2023, followed by Qwen-14B in September and Qwen-72B in December. The early models drew on the Llama architecture that Meta had popularized and used Alibaba's own Tongyi Qianwen license, which allowed commercial use below a large user threshold. Even so, Qwen-72B gave developers a strong open model with good Chinese and English at a time when few existed.
Qwen2 and Qwen2.5 (2024): The Developer Default
Qwen2, released in June 2024, moved most sizes to the Apache 2.0 license. Qwen2.5, released in September 2024, came in seven sizes from 0.5 billion to 72 billion parameters, with specialist Coder and Math versions. The range was the point. A developer could prototype on a small Qwen model on a laptop and move to a large one on a server without changing families.
By 2025, Qwen had become the most common base for community fine-tunes on Hugging Face. When DeepSeek released small distilled versions of R1, several of them were built on Qwen. Read the DeepSeek history for that side of the story.
QwQ (November 2024 to March 2025): Qwen Learns to Reason
QwQ-32B-Preview, released in November 2024 under Apache 2.0, was Qwen's first open reasoning model, built to think step by step before it answers. The full QwQ-32B followed in March 2025. A 32-billion-parameter model that could reason was small enough to run on a single high-end GPU. For background, read AI reasoning models explained.
Qwen3 (April 2025): Hybrid Thinking for Everyone
Qwen3, released on April 28, 2025, is the release most developers remember. It shipped under Apache 2.0 in eight sizes: six dense models from 0.6 billion to 32 billion parameters and two mixture-of-experts models, 30B-A3B and 235B-A22B. Alibaba trained it on about 36 trillion tokens covering 119 languages and dialects.
Qwen3 introduced hybrid thinking: one model could answer fast or think step by step, and the user chose. In July 2025 Alibaba split the two modes into separate Instruct and Thinking models again, and Qwen3-Coder, a 480-billion-parameter open model for agentic coding, followed the same month.
Qwen3-Max (September 2025): The Closed Flagship
Qwen3-Max, released in September 2025, has more than 1 trillion parameters and runs only as a paid service on Alibaba Cloud. It marked a clear line in the strategy. The open Qwen models would be strong and widely available, and the very top tier would be sold.
Qwen3.5 and Qwen3.6 (February to April 2026)
Qwen3.5, released on February 16, 2026, came as open weights for a 397-billion-parameter mixture-of-experts model, plus a proprietary 3.5-Plus. Qwen3.6, released in April 2026, added an open 35B mixture-of-experts model under Apache 2.0 and a proprietary 3.6-Plus. Both generations put more weight on agent work: tool calls, long tasks, and coding.
Qwen 3.7 Max and Plus (May to June 2026): Built for Agents
Alibaba announced Qwen 3.7 Max on May 20, 2026, at its cloud summit in Hangzhou, and built it for agentic coding and complex reasoning. It has a 1-million-token context window and targets long tasks that run for hundreds or thousands of steps, such as coding and office workflow automation. Qwen 3.7 Plus followed on June 1, 2026, at a lower price and with text, image, and video input. Alibaba did not release open weights for the 3.7 generation.
Qwen 3.8 (August 2026): A 2.4-Trillion-Parameter Open Model
Alibaba previewed Qwen 3.8 Max at the World AI Conference in Shanghai in July 2026 and launched it as a cloud model on August 3, 2026. On August 12, it released open weights for a 2.4-trillion-parameter mixture-of-experts model with about 95 billion active parameters per token. The open flagship weights use a custom license that adds terms for providers with more than $50 million in annual revenue. Qwen3.8-27B followed on August 14 under Apache 2.0, and Flash models followed later in August.
Is Qwen Open Source? Licenses Explained
Many Qwen models are open weight, but not all of them, and the license depends on the model. The rule of thumb in 2026: small and mid-size open models use Apache 2.0, the largest open flagship uses a custom license, and Max, Plus, and Flash tiers are usually API-only.
| Qwen release | License | What it means for you |
|---|---|---|
| Qwen3 (all sizes) | Apache 2.0 | Use, change, and sell freely |
| Qwen3.6 open 35B | Apache 2.0 | Use, change, and sell freely |
| Qwen3.8-27B | Apache 2.0 | Use, change, and sell freely |
| Qwen3.8 2.4T open flagship | Custom Qwen license | Extra terms for providers with more than $50M a year in revenue |
| Max, Plus, and most Flash tiers | Proprietary | API access on Alibaba Cloud and partners only |
As with every open-weight family, the training data and full training code are not released. Read the history of open-source LLMs for how these licenses evolved, or our guide to the best open-source LLMs.
Who Leads Qwen? From Lin Junyang to Alibaba Token Hub
Lin Junyang joined Alibaba in 2019 and became the technical lead of Qwen in 2023. Under him, the team grew Qwen from a single 7B model into a family with more than a billion downloads. Lin became the public face of Qwen on social media and at research conferences.
On March 4, 2026, shortly after the Qwen3.5 release, Lin announced that he was stepping down. Alibaba Cloud chief technology officer Zhou Jingren took interim charge of the Qwen team. On March 16, 2026, Alibaba moved Tongyi Lab, the Qwen unit, and its other AI teams into a new Alibaba Token Hub business group that CEO Eddie Wu leads directly. Alibaba said it stays committed to open source, and the open Qwen3.6 and Qwen3.8 releases that followed kept that promise.
Why Do So Many Developers Build on Qwen?
Developers build on Qwen because it offers the widest range of open sizes of any lab under a permissive license. In August 2026, Hugging Face counted more than 150,000 Qwen-based derivative models on its Hub, more than for any other model maker.
WHY QWEN BECAME THE DEFAULT BASE MODEL
------------------------------------------------------------
Range 0.6B on a phone -> 235B and up in a data center
License Apache 2.0 on most open sizes, commercial use OK
Languages 119 languages and dialects (Qwen3)
Variants Coder, Math, VL (vision), Omni (audio), Thinking
Cadence A new generation every few months, 2023 to 2026
------------------------------------------------------------
That cadence has a cost. A model you fine-tuned in spring can be two generations old by autumn. It is one reason teams increasingly pick models per task instead of committing to one forever.
Qwen vs Other Open-Weight Labs
Qwen shares the Chinese open-weight frontier with DeepSeek and Moonshot AI. Each lab leads on a different axis.
| Lab | Flagship family | Known for | Deep dive |
|---|---|---|---|
| Alibaba | Qwen 3.7 and 3.8 | Widest size range, agent work, Apache 2.0 | You are here |
| DeepSeek | V4 Pro, V4.1 Flash | Low cost, reasoning, MIT license | DeepSeek history |
| Moonshot AI | Kimi K3 | Long context, agent swarms | Moonshot and Kimi history |
For head-to-head detail, compare Qwen vs DeepSeek and Qwen vs Kimi.
How to Use Qwen Models Inside Taskade
Qwen 3.7 Plus is in the Taskade model picker on every paid plan, and Qwen 3.7 Max is available on Max and Enterprise. Taskade provides frontier models from top AI labs, from OpenAI GPT to open-weight models such as DeepSeek and Qwen, in one workspace. Claude is available on Enterprise with your own API key (BYOK).
Here is how to put Qwen to work:
- Leave TSK-Auto on. TSK-Auto is the default and picks a model for each turn.
- Pin Qwen 3.7 Plus for volume. On a custom AI agent, choose All Models and pick Qwen 3.7 Plus for agentic coding and long forms at a lower cost.
- Pin Qwen 3.7 Max for the hard jobs. On Max and Enterprise, use it as the flagship for long, multi-step agent work.
- Build an app with it. In the Taskade Genesis builder chat, choose a Qwen model from Custom Engine. The model guide explains each menu.
- Put agents on a team. Use multi-agent teams so one agent runs Qwen and another runs a different model, and hand the output to an automation.
A very large input can cost more per call on the Qwen family than its cost band suggests, so keep long inputs in mind at volume. Compare every family in the full model catalog.
How Does Qwen Do on Real App Builds?
The TSK-1 Benchmark tests models with apps, not quiz scores. Each test sends one fixed app request to every model family inside Taskade Genesis and checks the result the way a user would: does the app open, does it save every answer, and does a follow-up edit land?
The Qwen results page sums up the family as best for value on long, detailed forms, and it keeps a dated record of every test, the unfinished edits included. Read the method behind the tests, the first TSK-1 report, or see which model builds the best apps.
Pricing for Qwen in Taskade
You do not need an Alibaba Cloud account to use Qwen in Taskade. The AI allowance comes with your plan. You start free, then earn or buy AI credits as you build, and paid plans include a monthly credit allowance. Pro is $10 per month billed annually for up to 10 members, Business is $25, Max is $100, and Enterprise is $250, each per month billed annually. See pricing for every plan.
What's Next for Qwen?
Qwen's path through 2026 points to three things. Agent work is now the main target, from Qwen 3.7 Max's long tasks to Qwen3-Coder. Open weights keep growing, with the 2.4-trillion-parameter Qwen 3.8 release. And the release cadence stays fast, with a new generation every few months.
For builders, that cadence is the lesson. The best model for a job in October will not be the best in March. Keep the durable parts of your work above the model: the projects that hold your data, the agents that hold your instructions, and the automations that run your process. In Taskade, that is Workspace DNA. Projects remember ▲, agents think ■, and automations execute ●, whichever model runs underneath.
Related Reading
- DeepSeek History: the lab that made frontier AI cheap
- xAI and Grok History: from X to SpaceX
- Moonshot AI and Kimi History
- The History of AI Benchmarks
- AI Cost per Task
- How Large Language Models Work
Frequently Asked Questions About Qwen
What is Qwen?
Qwen is the family of large language models built by Alibaba Cloud, the cloud computing arm of Alibaba Group. The name is short for Tongyi Qianwen. Alibaba releases some Qwen models as open weights, most under the Apache 2.0 license, and keeps its largest Max, Plus, and Flash models as paid services on its cloud. Qwen is one of the most widely fine-tuned open model families in the world.
When was Qwen released?
Alibaba opened a beta of its Tongyi Qianwen chatbot in April 2023, released its first open-weight model, Qwen-7B, in August 2023, and opened the service to the public in September 2023 after regulatory approval. Major generations followed: Qwen2 in June 2024, Qwen2.5 in September 2024, Qwen3 in April 2025, Qwen3.5 in February 2026, and Qwen3.8 in August 2026.
What does Tongyi Qianwen mean?
Tongyi Qianwen is the Chinese name of Alibaba's model family. Tongyi is the brand Alibaba uses for its AI products, and Qianwen means a thousand questions, a nod to a model that can answer any question you ask it. Qwen is the short English name, built from the first letters of Qianwen.
Is Qwen open source?
Many Qwen models are open weight. Qwen3 shipped in April 2025 under the Apache 2.0 license in sizes from 0.6 billion to 235 billion parameters, and later releases such as Qwen3.6 and Qwen3.8-27B also use Apache 2.0. The largest models follow different rules. Max, Plus, and Flash tiers are usually API-only, and the open Qwen3.8 flagship weights use a custom license with extra terms for very large providers.
What is the difference between Qwen 3.7 Max and Qwen 3.7 Plus?
Qwen 3.7 Max, announced on May 20, 2026, is Alibaba's flagship model for agent work, with a 1-million-token context window and a focus on long, multi-step tasks. Qwen 3.7 Plus, released on June 1, 2026, is a lower-cost sibling that accepts text, image, and video input and suits agentic coding at volume. Both are proprietary and run on Alibaba's cloud.
What is Qwen 3.8?
Qwen 3.8 is Alibaba's newest model generation. Alibaba previewed Qwen 3.8 Max in July 2026, launched it as a cloud model on August 3, 2026, and released open weights for a 2.4-trillion-parameter version on August 12, 2026. A smaller Qwen3.8-27B followed on August 14 under Apache 2.0, and Flash models followed later in August.
Who leads the Qwen team?
Lin Junyang, the technical lead who grew Qwen from 2023, stepped down in March 2026. Alibaba Cloud chief technology officer Zhou Jingren took interim charge of the Qwen team. Later that month, Alibaba moved Tongyi Lab, the Qwen unit, and its other AI teams into a new Alibaba Token Hub business group that CEO Eddie Wu leads directly. Alibaba said after the change that it stays committed to open-source releases.
Why do so many developers build on Qwen?
Developers build on Qwen because Alibaba ships the widest range of open sizes of any lab, from models small enough for a phone to models that need a data center, mostly under the permissive Apache 2.0 license. That range makes Qwen a common base for fine-tuning, and in August 2026 Hugging Face counted more than 150,000 Qwen-based derivative models.
Can I use Qwen models inside Taskade?
Yes. Qwen 3.7 Plus is in the Taskade model picker on every paid plan, and Qwen 3.7 Max is available on Max and Enterprise. You can pin a Qwen model on an AI agent or in the Taskade Genesis builder chat, or leave TSK-Auto on and let Taskade pick a model for each turn. Paid plans start at Pro for $10 per month billed annually.
How does Qwen perform on real app builds?
The TSK-1 Benchmark sends the same app request to each model family inside Taskade Genesis and checks whether the app works, saves data, and accepts follow-up edits. The Qwen results page summarizes the family as best for value on long, detailed forms, and it keeps a dated record of every test, including the edits that did not finish.
Ready to build with Qwen and the rest of the frontier? Create your first app in Taskade Genesis →, browse AI agents and live community apps, or start from the Taskade homepage.