What is Qwen and how it can help you be more productive

Alibaba Qwen guide · Updated November 2026

Qwen isn’t just the open-weight family your local setup pulls from Hugging Face anymore. What began on April 11, 2023 as Alibaba Cloud’s Tongyi Qianwen enterprise beta has grown into one of the broadest model ecosystems in AI: open weights from a 27B dense model up to a 2.4-trillion-parameter MoE flagship, hosted Max and Flash tiers, a consumer app that passed 100 million monthly users within two months of launch, and workplace agents inside DingTalk. On November 11, 2026 Alibaba opened a Qwen 4 preview as qwen4-max-preview, and five days later, on November 16, Qwen4-Max went generally available: the first Qwen 4 flagship – a 2.4-trillion-parameter model with 71 billion parameters active per token – with text, image, and video input, a 1M-token context window, and a $1.50/$4.50 per 1M-token launch price that undercuts Qwen3.8-Max by 25%.

In this guide we explain what Qwen is in 2026, how the Qwen 4, Qwen 3.8, Qwen 3.5–3.7, and Qwen 3 families fit together, what Qwen Studio, QwenWork, Qoder, and Qwen Code can do, what the plans cost, and how to use it well – for research, coding, documents, and everyday productivity.

Apr 11, 2023Tongyi Qianwen unveiled by Alibaba Cloud
Qwen4-Max · Nov 16, 20262.4T backbone · 71B active · 1M-token context
Qwen Studio · QwenWork · QoderToken Plan from $6/mo · open weights included
01 · What it is

What is Qwen?

Qwen – known in China as Tongyi Qianwen, a name roughly meaning “truth from a thousand questions” – is Alibaba Cloud’s AI model family. Alibaba unveiled it on April 11, 2023 as an enterprise beta, opened it to the public that September, and has released open weights almost continuously since. The strategy is deliberately two-track: small and mid-size models ship as open weights, mostly under Apache 2.0, while the top Max models are hosted products – although Qwen3.8 broke the pattern by publishing the weights of its 2.4-trillion-parameter flagship. The scale is unmatched among open families: Alibaba reports more than 3 billion cumulative downloads, over 460 open models, and some 300,000 community derivatives, and Hugging Face alone counted 2.045 billion Qwen downloads in 2026.

The products reach further than most open families too. Qwen Studio (chat.qwen.ai, renamed from Qwen Chat on April 10, 2026) is the free assistant with Deep Research, Web Dev, and Slides modes. The Qwen app passed 100 million monthly users within two months of its November 2025 beta and now handles Taobao shopping, Fliggy travel, and Alipay payments in China. QwenWork bundles Alibaba’s workplace agents and passed 30 million users within a month of its August 2026 beta. Qoder and the open-source Qwen Code CLI serve developers, and Singapore’s Sea-Lion project even moved its base model from Llama to Qwen.

Today, depending on the surface, Qwen can:

  • Answer questions & explain concepts
  • Write, rewrite, summarize & translate (119 languages)
  • Search the live web with citations
  • Reason with adjustable thinking effort
  • Write, debug & ship code (Qwen Code, Qoder)
  • Understand images, documents & video
  • Generate images with rendered text (Qwen-Image)
  • Transcribe & generate speech (Qwen3-TTS/ASR)
  • Run workplace agents (QwenWork)
  • Shop, book travel & pay in China
  • Run locally on open weights
Qwen – the models

Numbered generations (Qwen, Qwen2/2.5, Qwen3, Qwen3.5–3.8, and now Qwen 4) in tiers: Max for flagship capability, Plus for balance, Flash for speed and cost, with specialists for coding, vision, speech, and images. Open weights cover everything from a 27B dense model to the 2.4T-A95B MoE; the newest Qwen4-Max flagship is hosted only.

Qwen – the products

Qwen Studio for free chat and research, the Qwen app for consumer and China errands, QwenWork for workplace agents, Qoder and Qwen Code for coding, Alibaba Cloud Model Studio and QwenCloud for the API, and Hugging Face and ModelScope for open weights.

02 · The model family

Qwen4-Max, Qwen3.8-Max, and the open-weight line

Qwen ships quickly and keeps several generations alive at once, which can be confusing. The current picture is one hosted flagship from the new Qwen 4 generation, one open-weight 2.4T flagship from Qwen 3.8, a cheap Flash tier, and a 27B dense model that runs on a workstation – plus the Apache 2.0 families many teams still run.

New flagship · Nov 16, 2026

Qwen4-Max

The first flagship of the Qwen 4 generation: a rebuilt sparse mixture-of-experts architecture with Gated DeltaNet linear attention, Qwen Sparse Attention, Gated Residual connections, and N-gram Embedding. Alibaba publishes it at 2.4 trillion backbone parameters with 71 billion active per token plus a 328-billion N-gram Embedding table. Text, image, and video input, a 1M-token context, thinking at low, high (default), or xhigh effort – and from this generation on, max_tokens counts thinking and the answer together. On Alibaba’s Work360 benchmark of industry tasks across 26 domains it scores 52.69 against 47.74 for Qwen3.8-Max. Hosted only: no weights, and no fine-tuning, batch, or dedicated deployments. The qwen4-max-preview ID opened on November 11 and retires on November 25.

Best for hard agentic & coding workQwen 4
Open-weight flagship · Aug 3 & 12, 2026

Qwen3.8-Max & 2.4T-A95B

The stable workhorse: a 2.4-trillion-parameter MoE with about 95 billion active parameters and a 1M-token context at $2/$6 per 1M tokens. On August 12 Alibaba published the weights as Qwen3.8-2.4T-A95B, the family’s largest open release, under the Qwen3.8-Max License – providers earning over $50 million a year on the model need a commercial agreement. The open checkpoint is text-only with thinking always on; the hosted service adds image and video input and a non-thinking mode.

Production default & largest open weights
Efficiency tier · Aug 26, 2026

Qwen3.8-Flash & Flash-Next

Flash-Next is the open-weight prototype of the Qwen 4 architecture: a 125-billion-parameter MoE with only 6 billion active plus a 51-billion n-gram embedding table, under the Qwen Community 1.0 licence. Qwen3.8-Flash is its production twin: the fast multimodal agent tier at $0.15/$0.47 per 1M tokens with a 1M-token context – roughly a twelfth of Max’s price for high-volume work.

High-volume agents at a fraction of MaxEfficiency
Self-hosting pick · Aug 14, 2026

Qwen3.8-27B

A dense 27-billion-parameter multimodal model under Apache 2.0, the most permissive licence in the Qwen 3.8 family: image input, thinking on by default, and a 262,144-token context. Alibaba says it beats Qwen3.7-Plus on coding and office tasks, and quantized it runs on a single workstation GPU.

One GPU with quantization & full control
Omni-modal · Sep 18, 2026

Qwen3.8-Omni-Flash

Text, image, audio, and video in, text out, with agentic tool use and a 1M-token context. Built for perception-plus-action loops: transcribe the meeting, read the slides, then draft the summary in one session.

Perception plus tool execution

Around the core sit the previous generations and the specialists: Qwen3 (April 28, 2025 – Apache 2.0, hybrid thinking, 0.6B to 235B-A22B, 119 languages), Qwen3.5 (February 16, 2026 – the open 397B-A17B flagship plus a hosted Plus), Qwen3.6 and the Qwen3.7 Max and Plus line, Qwen3-Coder (480B-A35B for repo-scale work), Qwen3-VL for vision-language, and the media stack – Qwen-Image 3.0, Wan 2.7, HappyHorse 1.1 video, Qwen-Audio 3.0, plus Qwen3-TTS and Qwen3-ASR for speech. Alibaba says smaller open-weight Qwen 4 models will follow, without a date.

03 · How it works

How does Qwen work?

Qwen’s flagships are sparse mixture-of-experts transformers, and for Qwen4-Max Alibaba describes three upgrades in the new architecture. Hybrid attention mixes Gated DeltaNet linear-attention layers with Qwen Sparse Attention, which picks relevant micro-blocks of context instead of scanning every token; Gated Residual connections with learned read and write gates keep training stable at this scale; and N-gram Embeddings add a large extra parameter table that increases capacity with little extra compute. Thinking is on by default at high effort, with xhigh available for the hardest tasks and low for routine work.

A typical request flows through several stages:

  • 1

    Understand the request

    The model reads your prompt, the conversation history, and anything you uploaded – files, images, video, code – plus tool definitions, connected MCP servers, and cached context.

  • 2

    Reason between steps

    Thinking runs at high effort by default; escalate to xhigh when a task fails, drop to low for routine work. Alibaba reports token usage dropping by 24–35% on Work360 tasks and by roughly half on Terminal-Bench 4.0, with the average cost per task falling from $4.43 to $1.67.

  • 3

    Decide which tools to use

    Built-in web search and a code interpreter, function calls, or a coding harness such as Qwen Code, Qoder, Claude Code, or Cursor pointed at Alibaba’s OpenAI- and Anthropic-compatible endpoints.

  • 4

    Generate the response or take the action

    An answer, a patch, a slide deck, a website, or a multi-step workflow carried out in QwenWork – with Live Delivery shipping a usable first version early and improving it while you steer.

  • 5

    Keep the session coherent

    A 1M-token window, context caching at $0.075 per 1M cached tokens, and – in QwenWork – optional memory and reusable skills saved from workflows that worked.

One practical trap in the new generation: from Qwen4-Max on, max_tokens counts thinking and the answer together, so a limit that is too low can leave you with no answer at all. The combined thinking-plus-answer ceiling is 262,144 tokens – leave room for reasoning in your budget, especially at xhigh effort.

04 · What you can do

What can you do with Qwen?

Research and information

Qwen Studio (chat.qwen.ai, renamed from Qwen Chat on April 10, 2026) is free on web, mobile, and desktop, with web search, Deep Research, Slides, and Translation modes plus a code interpreter. It makes a useful second opinion alongside your main assistant.

Writing and translation

Draft, rewrite, summarize, and translate – Qwen3 introduced support for 119 languages and dialects, and Chinese-English work is a particular strength. Keep files and instructions together so every session starts with the right context.

Coding and DevOps

Qwen Code and Qoder plan, edit, and test across a repository, and Qwen4-Max adds Live Delivery: the agent ships a usable first version early and improves it release by release while you steer. Qwen3-Coder-480B-A35B is tuned for repo-scale agentic work.

Long-document analysis

A 1M-token window holds an entire codebase, book-length contract sets, or hours of transcripts in one prompt. Test retrieval on your own documents before trusting the middle of a very long context.

Data analysis

Upload CSV, XLSX, and PDFs in Qwen Studio; the models clean data, compute metrics, find trends and outliers, and render charts. For bigger jobs, QwenWork turns research into an actual artifact: a comparison spreadsheet, a budget, or a weekly briefing.

Images and visual analysis

Multimodal models read screenshots, diagrams, photos, and documents, and Qwen-Image generates and edits images with complex text rendering inside them – where most image models still fail. Wan 2.7 and HappyHorse 1.1 cover image and video generation.

Voice and conversation

Qwen3-TTS generates stable, expressive, and streaming speech with custom voices, Qwen3-ASR transcribes multilingual audio for downstream pipelines, and Qwen3.8-Omni-Flash takes audio and video in directly alongside text.

Apps, integrations, and self-hosting

The API speaks OpenAI- and Anthropic-compatible protocols, so Claude Code, Cursor, Codex, and Qwen Code all work through one Token Plan. Open weights from Hugging Face and ModelScope run on vLLM and SGLang for self-hosting.

05 · From chatbot to agent

The agentic shift: Qwen Studio, QwenWork, Qoder, and Qwen Code

The biggest change since 2023 is not that answers got better, it is that Qwen moved from conversation toward execution: you define a goal, Qwen researches, uses tools, and comes back with something finished. Alibaba now sells one intelligence across several surfaces, and picking the right one for the job is most of the productivity win.

01

Qwen Studio – the free assistant

The zero-setup route at chat.qwen.ai: sign in, open the model menu, and pick Qwen4-Max. Free with Deep Research, Web Dev, Slides, and Translation modes plus a code interpreter and image generation. Alibaba has not published a Qwen4-Max quota for it, so expect limits on heavy use.

Free · web, mobile, and desktop

02

QwenWork – the workplace agent

QwenWork bundles Alibaba’s QoderWork, MuleRun, and Wukong agents into one workspace that drafts documents, builds websites, and analyzes data, with optional memory and reusable skills saved from workflows that worked. It passed 30 million users within a month of the August 2026 beta, Jefferies ranked it first among eight workplace agents, and it sits inside DingTalk in China.

Free plan · QoderWork + MuleRun + Wukong

03

Qoder, Qwen Code, and the API

Qoder is Alibaba’s agentic coding platform (IDE, CLI, and JetBrains plugin) with early access to new models. Qwen Code is the open-source terminal agent adapted from Google’s Gemini CLI – Apache 2.0, 27,000-plus GitHub stars, and 68.7 agent readiness in the Anchor Terminal comparison cited in our MiniMax guide; the tool is free and you pay for the model it calls. Underneath, the QwenCloud API serves pay-as-you-go tokens from Singapore, Frankfurt, Virginia, Tokyo, Hong Kong, and Beijing.

Qoder · Qwen Code · pay-as-you-go API

Qwen4-Max’s Live Delivery is the pattern to watch: not one perfect answer, but a usable first version shipped early and improved release by release while you steer. And in China the same intelligence also shops: since May 11, 2026 the Qwen app searches Taobao and Tmall’s four-billion-plus products, books travel through Fliggy, navigates with Amap, and pays through Alipay inside the chat.

06 · Key moments

The Qwen timeline: 2023–2026

Apr 11, 2023

Alibaba Cloud unveils Tongyi Qianwen as an enterprise beta for its cloud customers.

Aug 2023

The first open weights arrive as Qwen-7B, followed by Qwen-VL for vision-language work.

Jun 2024

Qwen 2 ships in sizes from 0.5B to 72B, including a 57B-A14B mixture-of-experts variant.

Sep 2024

Qwen2.5 lands with more than 100 open models plus dedicated Coder and Math lines, and becomes a common fine-tuning base.

Apr 28, 2025

Qwen 3 arrives under Apache 2.0 with hybrid thinking and a family from 0.6B to 235B-A22B, trained in 119 languages.

Jul 2025

Qwen3-Coder-480B-A35B launches for repo-scale agentic coding, together with the open-source Qwen Code CLI.

Sep 2025

Qwen3-Max passes a trillion parameters; Qwen3-Omni and Qwen3-VL extend the family to audio and vision.

Nov 2025

The Qwen app opens in public beta and passes 100 million monthly users within two months.

Feb 16, 2026

Qwen3.5 ships, starting with the open 397B-A17B flagship plus a hosted Plus tier.

Apr 10, 2026

Qwen Chat is renamed Qwen Studio at the same address, chat.qwen.ai.

Apr–Jun 2026

Qwen3.6 and the Qwen3.7-Max and Plus models follow; Qwen3.7-Plus launches at $0.40/$1.60 per 1M tokens.

May 11, 2026

The Qwen app gains Taobao and Tmall shopping, Fliggy travel, Amap navigation, and Alipay payments in China.

Jul 19, 2026

Qwen3.8-Max previews at WAIC Shanghai ahead of its August launch.

Aug 3, 2026

Qwen3.8-Max goes GA: a 2.4-trillion-parameter MoE with 95B active parameters, launched alongside the QwenWork beta, which passes 30 million users within a month.

Aug 12, 2026

Open weights publish as Qwen3.8-2.4T-A95B, the family’s largest open release, under the Qwen3.8-Max License.

Aug 14, 2026

Qwen3.8-27B ships as a dense Apache 2.0 multimodal model with a 262,144-token context.

Aug 26, 2026

Qwen3.8-Flash and Flash-Next ship, Flash-Next doubling as the public preview of the Qwen 4 architecture; QwenWork International opens.

Sep 18, 2026

Qwen3.8-Omni-Flash ships for text, image, audio, and video understanding with agentic tool use.

Nov 11, 2026

The Qwen 4 preview opens as qwen4-max-preview, available on Token Plan and in Qwen Studio, the Qwen app, QwenWork, and Qoder.

Nov 16, 2026

Qwen4-Max reaches general availability with a public pay-as-you-go API at $1.50/$4.50 per 1M tokens.

07 · Staying productive

How to use Qwen well

Match the model to the job

Qwen4-Max for hard agentic and long-horizon work, Qwen3.8-Flash for high-volume drafting and extraction at roughly a twelfth of the price, and the open 27B when data has to stay on your own hardware.

Pin the exact model ID

qwen4-max, qwen4-max-preview, and dated snapshots such as qwen3.8-max-0902 behave differently and have different limits and retirement dates – the preview ID, for example, retires on November 25. Record the model ID, region, and thinking effort with every evaluation.

Use the cache and the default effort

Cached input costs $0.075 per 1M tokens against $1.50 uncached, so put stable instructions and documents first in the prompt. Keep thinking at the default high effort and escalate to xhigh only when a task fails.

The prompt pattern that works

Role:You are a senior engineer working in this repository.Task:Find the failing test, fix the implementation, and run the relevant tests. Do not change the public API.Format:A short plan first, then the changes as bullets with file and line references, and the test output at the end.

Role, task, format – the same structure powers everything from pull-request reviews to research briefs. For agents, add the contract: the spec, the tests, the definition of done, and what needs your approval before anything ships or spends.

08 · The window

How large is the context window?

The flagship tier now carries a 1,000,000-token context window – roughly 750,000 words – enough to hold whole codebases, book-length contracts, or an entire project history in a single session. For Qwen4-Max, Alibaba documents up to 983,616 input tokens in thinking mode (991,808 without thinking), with output capped at 131,072 tokens without thinking and 262,144 for thinking plus the final answer combined.

Qwen4-Max / Qwen3.8-Max / Qwen3.8-Flash – 1,000,000 tokens≈ 750,000 words
Qwen3.8-27B / Flash-Next – 262,144 tokens≈ 196,000 words
Qwen 2.5 – 128,000 tokens≈ 96,000 words

Practical rule of thumb: a token is roughly 0.75 English words, and about 1.5 tokens per word once you factor in punctuation and spacing. Thinking, tool output, and the answer all draw from the same window, and attention is not perfectly uniform across it – expect some “lost in the middle” degradation on very long documents, and test retrieval on your own material before relying on it.

09 · Plans and pricing

What does Qwen cost?

PlanPriceModel accessLimits
Qwen Studio$0Qwen chat at chat.qwen.ai with Deep Research, Web Dev, Slides, and TranslationFree to start; Qwen4-Max quotas not published
QwenWorkFree plan availableQoderWork, MuleRun, and Wukong agents for documents, websites, and dataEnterprise connectors planned at the international launch
Token Plan Lite$6/mo ($3 for new users)20,000 credits for Qwen Code, Qoder, Claude Code, Cursor, and CodexSingapore region; third-party models such as DeepSeek bundled
Token Plan Essential$10/mo ($5 for new users)45,000 credits, same model accessHigher concurrent-agent tier
Token Plan Standard$18/mo80,000 creditsHigher quota
Token Plan Pro$43/mo200,000 creditsHighest individual quota
Token Plan Team≈$20–$200/seat/moSame models, seat management, usage analyticsMonthly credit pool per seat; no training on your data
API · qwen4-max$1.50 / $4.50per 1M tokens in/out, 1M context, text, image, and video input$0.075 cached input · thinking low/high/xhigh · no fine-tuning, batch, or dedicated hosting
API · qwen3.8-max$2.00 / $6.00per 1M tokens in/out, 1M context$0.25 cached input · weights also published as 2.4T-A95B
API · qwen3.8-flash$0.15 / $0.47per 1M tokens in/out, 1M contextHigh-volume agent tier
Open weightsFree (licence)Qwen3.8-27B, Flash-Next, 2.4T-A95B, and earlier families from Hugging Face and ModelScopeHardware is on you; licences differ by model

Qwen4-Max is 25% cheaper on input and output than Qwen3.8-Max and 70% cheaper on cached input ($0.075 against $0.25) – the cache rate matters most for agents, because a coding agent resends the conversation, tool schemas, and file contents on every turn. Token Plan figures are the November 2026 tiers, credits are a bundled unit without a published token conversion, and new-account token promotions appear in some regions. The table reflects November 16, 2026 – check Alibaba’s official QwenCloud and Model Studio pages for current rates.

10 · FAQ

Frequently asked questions

Is Qwen free?

Yes to start. Qwen Studio is free at chat.qwen.ai with Deep Research and creation tools, QwenWork has a free plan, Token Plan subscriptions start at $6 a month for coding-tool usage, and most open-weight models cost nothing but the hardware to run them. API use is pay per token.

Is Qwen open source?

Mostly, with exceptions. Qwen3.8-27B, the Qwen 3 family, and Qwen3.6’s 35B-A3B are Apache 2.0; Flash-Next uses the Qwen Community 1.0 licence; and the 2.4T Qwen3.8-Max weights require a commercial agreement from providers earning over $50 million a year on the model. Qwen4-Max is hosted only today – Alibaba says smaller open Qwen 4 models will come later, without a date.

What is qwen4-max-preview?

The preview model ID Alibaba opened on November 11, 2026, the first public look at the Qwen 4 generation. It is Token Plan-only and also appears in Qwen Studio, the Qwen app, QwenWork, and Qoder. It retires on November 25, 2026, when the stable qwen4-max ID takes over.

Is Qwen better than ChatGPT, Claude, or Gemini?

It depends on the job. Qwen wins on open-weight choice, multilingual coverage, and price per task – Artificial Analysis measured $3.27 per task for Qwen4-Max against $10.02 for Qwen3.8-Max, while Kimi K3.1 Pro still leads Terminal-Bench 4.0 at 67% against Qwen’s 48.5%. Test on your own tasks, not on a leaderboard.

Can I run Qwen locally?

Yes for most sizes. The 27B dense model runs on a single workstation GPU with quantization, Flash-Next has community builds, and the 2.4T-A95B weights need a multi-GPU cluster; Qwen4-Max itself cannot be self-hosted. Use vLLM or SGLang, and read each licence first.

Is Qwen safe for business data?

Treat it like any cloud AI: check your contract, data residency, and retention settings before sending anything confidential. The API runs in Singapore, Frankfurt, Virginia, Tokyo, Hong Kong, and Beijing, Token Plan Team states your data is not used for training, and self-hosted open weights keep data inside your own boundary entirely.

How is Qwen different from DeepSeek, Kimi, or MiniMax?

DeepSeek leans on cheap open weights, Kimi on long context and agent swarms, MiniMax on multimodal output and coding price. Qwen leans on range and distribution: sizes from a workstation-friendly 27B to a 2.4T flagship, a cloud to run them on, a consumer app with hundreds of millions of users, and workplace agents inside DingTalk.

Heads-up:Qwen is moving unusually fast. The Qwen4-Max preview opened on November 11, 2026, general availability followed on November 16, and the preview model ID retires after November 25; smaller open Qwen 4 models are promised without a date, and Token Plan tiers changed several times this year. The details above reflect November 2026 – check Alibaba’s official Qwen and QwenCloud pages for current pricing, limits, and availability.

Related reading