QWN
// LLMs · language models

Qwen

Alibaba’s AI in three formats: a free chat, open weights for your own server and a pay-as-you-go API.

Vendor · Alibaba Cloud ↗
Pricing
Free chat · API Qwen3.8-Flash at $0.15 per 1M input tokens and $0.47 output (Singapore list price) · prices in USD
Model
Usage-based
Evaluation
External evaluation
Verified
November 16, 2026
◆ Operator verdict

Qwen is Alibaba’s family of AI models, and today you can use it three ways: the free Qwen Studio chat, the open weights (Qwen3.8-27B under Apache 2.0) and the QwenCloud API, where Qwen3.8-Flash costs $0.15 per million input tokens in Singapore. I build open models and APIs like Qwen into agents and automations when cost per token or running the model on your own server is what makes or breaks the project, and I train the team that will maintain it. What the homepage won’t tell you: the large models aren’t Apache 2.0, the free API quota only exists in Singapore, and the new Qwen4-Max, which reached the API on November 16, 2026, has no open weights.

+ What it’s GOOD for

  • Agents and automations that generate a lot of output: Qwen3.8-Flash charges $0.47 per million output tokens and has no peak-hour rate
  • Companies that want the model on their own server: Qwen3.8-27B is a free download under the Apache 2.0 license
  • Very long documents and conversations: the Qwen3.8 API models support a 1-million-token context window
  • Teams already using OpenAI- or Anthropic-compatible tools (Claude Code, Codex, Qwen Code): you just swap the base URL and the key
  • Text, images and video in a single model: Qwen3.8-27B and the API’s Flash and Max models understand images and video natively
  • Fine-tuning on your own data with small open models, such as Qwen3.5 from 0.8B to 9B

− What it’s NOT for

  • Customer personal data or regulated industries without a review first: I couldn’t verify the data retention policy for the pay-as-you-go API or the chat
  • Building an AI service for third parties on top of the large models: Qwen3.8-2.4T-A95B and Qwen3.8-Flash-Next ship under custom licenses with conditions
  • Anyone who wants Qwen4 on their own server: Qwen4-Max is only available through the API or Token Plan, and there are no open Qwen4 weights
  • Budgets that need a fixed price in a currency other than dollars: QwenCloud bills in USD in every region
  • Anyone assuming their data stays in the EU in Frankfurt: the default workspace uses the Global scope, which can run inference on any node, mainland China included
  • Teams that want everything outside Singapore: the free quota and Token Plan only work there, and fine-tuning and batch processing only exist in Singapore and Beijing

◇ Where it fits across StratoAI lines

AI strategist at night in front of a monitor showing a diagram of a model family branching into small, medium and large sizes, next to a terminal and a workstation under the desk

Qwen is Alibaba’s family of AI models: some you can download for free and run wherever you like, others are only available through a paid API, and you can try the current generation in a free chat. That generation is Qwen3.8, released in August 2026, and its fast API model, Qwen3.8-Flash, costs $0.15 per million input tokens in Singapore. Since November 16, 2026, the API also offers Qwen4-Max, the first Qwen4 model. I build open models and APIs like Qwen into business agents and automations, and I teach teams how to maintain them. Here’s what each piece is, what it costs and when I wouldn’t recommend it.

What Qwen is today (and which name means what)

Qwen (Tongyi Qianwen in China) is built by Alibaba Cloud’s Qwen team. The confusion on Google comes from the same name being used for models, a chat app, an API platform and several tools. Here’s the map:

NameWhat it isWhere
QwenThe model family: text, code, vision, speech and imagesThis page
Qwen StudioThe official chat app, free (formerly “Qwen Chat”). Its site also links to the official mobile appchat.qwen.ai
QwenCloudThe official API platform, with OpenAI- and Anthropic-compatible endpointsqwencloud.com
Token PlanA QwenCloud credits subscription for using the models from Claude Code, Cursor, Codex, Qwen Code and other toolsQwenCloud
Qwen CodeAn open-source coding agent for the terminalqwen.ai/qwencode
Qoder and QwenWorkAlibaba’s agentic coding platform and its AI work platformNot covered here
Qwen-ImageThe image model line (Qwen-Image-2.1 came out in September 2026)Not covered here

Watch out for third-party apps: one called “Q AI” shows up in Google’s results for “qué es qwen” (“what is qwen” in Spanish), and it isn’t the official one; the official Android app is listed on Google Play under “Alibaba Qwen Team.” If you were looking for the old API under the name “DashScope,” it’s now sold as QwenCloud, although the DashScope protocol is still one of its endpoints.

The current generation: Qwen3.8

Qwen3.8 arrived in August 2026. The official README puts it this way: “For the first time, Qwen3.8 brings a Qwen-Max-class model to open release.” According to the same README, it improves on coding, professional work, research and long-horizon agentic tasks, and it lets you tune how much the model reasons (reasoning_effort).

ModelTypeLicenseContextWhere you use it
Qwen3.8-27BDense, 27 billion parameters; understands text, images and videoApache 2.0262,144 tokens natively, extendable to 1,000,000Weights on Hugging Face and ModelScope; also via API
Qwen3.8-2.4T-A95BMoE: 2.4 trillion total parameters, 95 billion active, 512 experts; text only, and it always reasons before answering“Qwen3.8-Max License” (custom)262,144 natively, extendable to 1,010,000Weights on Hugging Face and ModelScope; also via API ($2 and $6 in Singapore)
Qwen3.8-MaxOfficial hosted version built on the 2.4T, with vision, built-in tools and a non-thinking modeAPI only1 millionQwenCloud API
Qwen3.8-Flash-NextExperimental MoE: 125 billion total, 6 billion activeQwen Community License 1.0262,144 natively, extendable to 1,000,000Weights on Hugging Face
Qwen3.8-FlashOfficial hosted version built on Flash-Next, fast and cheapAPI only1 millionQwenCloud API

Flash-Next comes with an important caveat: its official model card calls it an “experimental preview of the architecture that will underpin Qwen4.” In other words, it’s for trying out what’s coming, not a stable foundation for a product.

The earlier generations are still around. Qwen3.5 (February–March 2026) brought small models that are handy for fine-tuning on your own data: 0.8B, 2B, 4B and 9B, plus 27B, 35B-A3B, 122B-A10B and 397B-A17B. Qwen3.6 (April 2026) added a 35B-A3B and a 27B. All of them are Apache 2.0. Qwen3.7 (May–July 2026) was API-only. If you see articles mentioning a “Qwen 3 7B,” that size doesn’t exist: the dense models in the original Qwen3 series were 0.6B, 1.7B, 4B, 8B, 14B and 32B.

◆ Licenses: not all Qwen is equally “open source.” Apache 2.0 covers Qwen3.5, Qwen3.6, Qwen3.8-27B, Qwen3-Coder-Next and the speech models. The two large 3.8 models have their own licenses. The 2.4T’s “Qwen3.8-Max License” requires you to display the model name if your commercial product passes 100 million monthly active users or $20 million in monthly revenue. On top of that, if your company (together with its affiliates) runs a Model as a Service or AI Work Assistant business and has more than $50 million in total revenue over any 12 consecutive months, it needs a separate license from Qwen for any commercial use; internal use is exempt as long as you don’t make the model or its outputs available to third parties. Flash-Next’s Qwen Community License 1.0 has the same two conditions, but the separate license applies to any business of that kind, with no revenue threshold. Before you build a product on top of one, read that specific model’s LICENSE file.

How to use Qwen: three paths

1. Free chat. Qwen Studio is the official chat app. The Qwen3.8 README describes it as an assistant that’s “free to use, open to all.” It’s the fastest way to see how it responds before you touch anything technical.

2. The QwenCloud API. Pay as you go, with OpenAI-compatible (Chat Completions and Responses) and Anthropic-compatible (Messages) endpoints, so most tools and libraries work once you change the base URL and the key. There are six regions: Singapore, Beijing, Frankfurt, Hong Kong, Tokyo and Virginia (US). Each one has its own keys, and a key from one region won’t work in another. If you work in Claude Code, Cursor, Codex or Qwen Code itself, Token Plan gives you a credits subscription with its own endpoint instead of paying per token.

3. Open weights. You can download the open models from Hugging Face and ModelScope, and the official README lists Transformers, vLLM, SGLang, llama.cpp (GGUF format) and MLX (Apple silicon) for running them. For fine-tuning on your own data it recommends Unsloth, Swift or LLaMA-Factory. If you’d rather use a desktop app with no terminal, my LM Studio review explains how it works; as of November 16, 2026, Qwen3.8-27B was listed in its official catalog as a local download. And if you want to try Qwen alongside other models with a single key, OpenRouter has a Qwen page; I haven’t checked which models it serves today or at what price.

Open notebook on a wooden desk with a hand-drawn diagram that splits into two paths, toward an in-house server and toward the cloud, next to a pair of glasses, a fountain pen and a monitor showing code
The first decision with Qwen: weights on your own server or an API in the cloud.

Beyond chat. The family includes dedicated models that people search for separately: Qwen-Image for images, Qwen3-Coder-Next (January 2026, Apache 2.0) for code, and Qwen3-TTS and Qwen3-ASR for text-to-speech and speech-to-text (January 2026, Apache 2.0). On languages, Alibaba says that since Qwen3.5 it covers 201 languages and dialects; I haven’t measured its quality myself.

Qwen API pricing today

Official prices from the QwenCloud pricing table, checked on November 16, 2026; the cached-input price comes from each model’s page on qwencloud.com, and qwen4-max’s from its cache docs. These are list prices in dollars per million tokens, Singapore region. The billing docs state that prices are listed and billed in USD in every region.

ModelInputOutputImplicit cached inputNotes
qwen3.8-flash$0.15$0.47$0.0161M context, multimodal
qwen3.8-27b$0.50$3.00$0.101M context (hosted)
qwen3.8-max$2.00$6.00$0.251M context; 131K max output
qwen4-max$1.50$4.50$0.075Since November 16, 2026; 1M context; text, images and video
qwen3.7-max$2.50$7.50—Previous generation
qwen4-max-preview———No per-token price: Token Plan only, until November 25, 2026
Hands writing columns of figures in a notebook next to a vintage mechanical calculator and a cup of coffee, with an out-of-focus monitor showing a bar chart
Before I pick a model, I work out the cost using the workflow’s real input and output tokens.

What you should know before you pay:

  • Prices vary by region and scope. Under the “Global” scope (Frankfurt, Hong Kong, Tokyo and Virginia), qwen3.8-flash drops to $0.113 input and $0.382 output, qwen3.8-max to $1.65 and $4.951, and qwen4-max to $1.413 and $4.24 (the same as in Beijing); the trade-off is that this scope can run inference on any node, mainland China included. Under Frankfurt’s EU-only scope, qwen3.8-flash costs the same as in Singapore ($0.15 and $0.47), and under Virginia’s US-only scope, qwen3.8-max costs $2 and $6.
  • Batch processing. Only available in Singapore and Beijing.
  • A free quota, with fine print. New accounts get a quota per model, usually 1 million tokens (input and output combined), valid for 90 days from activation. It only exists in Singapore, only covers real-time calls (not batch, built-in tools like web search, or fine-tuning), and you have to add payment details to unlock it. The QwenCloud homepage mentions “70M+ free tokens,” a marketing figure the docs don’t break down.
  • Token Plan. The individual subscription has four tiers, according to its docs: Lite ($6 a month), Essential ($10), Standard ($18) and Pro ($43). Lite and Essential have a new-user price of $3 and $5 a month until November 30, 2026. The per-seat Team edition, according to its docs, guarantees your conversation data isn’t used to train models. Token Plan is only available in Singapore.

◇ Taxes: not spelled out. QwenCloud bills in dollars in every region, and the pages I reviewed don’t say how sales tax or VAT is applied to your account. Check the final amount in your billing console before you budget for it.

Qwen4-Max: now on the API (and what happens to the preview)

On November 16, 2026, QwenCloud announced Qwen4-Max (qwen4-max) in its changelog: the first Qwen4 model you can use through the pay-as-you-go API. Its official page presents it as “the next-generation flagship multimodal model in the Qwen family.” Here’s what the docs say today:

  • Per-token price. In Singapore, $1.50 per million input tokens and $4.50 output; in Beijing and under the Global scope (Frankfurt, Hong Kong, Tokyo and Virginia), $1.413 and $4.24. According to the billing page, it costs the same with or without reasoning, with no length tiers. It’s cheaper than qwen3.8-max ($2 and $6 in Singapore).
  • Its own cache. It debuts an explicit cache just for Qwen4 that lasts 5 minutes or 1 hour. In Singapore, writing to it costs $1.875 (5 minutes) or $3 (1 hour) per million tokens, and reading from it $0.075; in Beijing and under the Global scope, $1.767, $2.827 and $0.071.
  • Context and modalities. A 1,000,000-token context window, with up to 131,072 output tokens without reasoning and 262,144 for reasoning plus the answer combined. It understands text, images and video, answers in text and works with Chat Completions, Responses, Anthropic Messages and DashScope.
  • Free quota. 1 million tokens, only on pay-as-you-go calls in Singapore and separate from Token Plan credits.
  • What it doesn’t have. Batch processing only exists in Beijing, and the docs are clear: “Fine-tuning and model deployment are not supported.”
  • No open weights. You can only use it through the API or Token Plan. As of November 16, 2026, there was no Qwen4 repository on Hugging Face or on Qwen’s GitHub. I couldn’t verify the claims going around about a “Qwen 4 27B.”

What about Qwen4-Max-Preview? The preview that appeared on November 11 is still a separate model (qwen4-max-preview), Token Plan only, and the Token Plan docs keep its expiration date: it ends on November 25, 2026 at 24:00 (China time, UTC+8), and “the model will be retired after the preview ends.” Token Plan now also includes the final qwen4-max. The docs say to use “the appropriate model ID for each product” and don’t say the preview turns into qwen4-max on its own, so I’m not assuming any automatic migration: if you use it, switch the model ID before the 25th.

My take: it’s a brand-new model and I haven’t read an independent evaluation I can cite. Before taking it to production, I’d test it against qwen3.8-max on your own data. And if you need the model on your own server, the open generation is still Qwen3.8.

Privacy, data and regions

This is what QwenCloud’s official documentation says, plus what I couldn’t verify:

  • The region decides where your data is stored. According to the regions guide, the region determines where requests are received and where data is stored.
  • Singapore processes outside mainland China. Its scope is “International”: nodes outside the Chinese mainland. Beijing processes in mainland China.
  • Frankfurt has two scopes, and the default isn’t the European one. The “EU” scope processes in the European Union only; the “Global” scope, used by the default workspace, can run on any available node, mainland China included. The scope is set when you create the workspace and can’t be changed later. In today’s pricing table, Frankfurt lists qwen3.8-flash under both scopes (EU and Global) and qwen3.8-max and qwen4-max under Global only.
  • Training on your data. Token Plan Team rules out in writing that your conversation data is used to train models. I couldn’t verify the retention policy for the pay-as-you-go API or for the Qwen Studio chat in this review.
  • Weights on your own server. If you run an open model on your own infrastructure, your data never leaves it. That’s the route I suggest when sensitive data is involved.

If you’ll be handling customers’ personal data or you work in a regulated industry, get the guarantees in writing before you send anything to the API.

Qwen alternatives

Entry prices from each vendor’s official pages, in dollars per million tokens, checked on November 16, 2026.

AlternativeReference modelInputOutputNote
Qwen (baseline)qwen3.8-flash$0.15$0.47Singapore; no peak-hour rate
DeepSeekdeepseek-flash (V4.1-Flash)$0.15 off-peak · $0.30 peak$0.60 · $1.20Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. deepseek-v4-pro: $0.66 and $1.98 off-peak. DeepSeek review
Kimi (Moonshot AI)kimi-k3.1-pro$2.00$10.00kimi-k3: $3 and $15. The K2 series is scheduled to be retired at the end of December 2026. Kimi review
MistralMistral Small 4 · Mistral Large 4$0.15 · $0.70$0.60 · $2.10The European option. Large 4 per its official announcement; Mistral’s pricing page uses “Mistral Large” at $0.50 and $1.50 as an example, so confirm before you budget. Mistral review
Meta (Muse, formerly Llama)Muse Spark Open (open weights)Not verifiedNot verifiedllama.com now redirects to Meta’s developer site, which bills Muse Spark Open as “The most capable US open-weight model to date.” Llama review

Is Qwen or DeepSeek cheaper? On the fast models they tie on input ($0.15), but Qwen3.8-Flash is cheaper on output ($0.47 versus DeepSeek’s $0.60 off-peak) and doesn’t go up at peak hours, when DeepSeek doubles its rate. On the large models, qwen3.8-max ($2 and $6) is cheaper on output than kimi-k3.1-pro ($2 and $10), and qwen4-max ($1.50 and $4.50) is cheaper still. I won’t give you a verdict on quality: I haven’t read an independent comparison I can cite, so my advice is to test both on your own data.

And against ChatGPT or Claude? They’re closed models with their own ecosystems and their own pricing; you’ll find them in the directory (OpenAI and Claude). The useful question isn’t which one is “better,” but what task, what volume and what data you have.

My verdict as an operator

Qwen lets you choose between downloading the model and paying per use without switching vendors: a 27B under Apache 2.0 you can run on your own server, an API whose fast model costs $0.15 per million input tokens, and speech, code and image models around it. For agents and automations that generate a lot of text, Qwen3.8-Flash’s output price works in its favor.

What I won’t hide from you: the large models come with conditional licenses, the free quota and Token Plan only live in Singapore, in Frankfurt the default scope isn’t the European one, and I couldn’t verify its data policy. My recommendation: start with the chat, measure one real workflow on the Flash API and, if sensitive data is involved, move to open weights on your own infrastructure. And if you want to try Qwen4, do it with qwen4-max on the API: the Token Plan preview expires on November 25.

I’ll build it in and teach you how to use it

I build open models and APIs like Qwen into business agents and automations, and I train the teams who’ll maintain them. The work comes down to picking the model and region that fit your data, costing it out against your real volume and leaving it running without you depending on me.

  • If you want an agent built on the right model: AI agent implementation.
  • If you want the model wired into your processes: Automations.
  • If you want me to teach you and your team to work with open models and APIs: AI Mentoring.
  • If you’re not sure where to start: tell me about your situation and I’ll tell you whether Qwen is a fit or whether something else makes more sense.

Qwen FAQ

What is Qwen?

Qwen is Alibaba’s family of AI models. It handles chat, writing, coding, long-document analysis, image and video understanding, image generation, and speech-to-text and text-to-speech. Some models ship as open weights and others are API-only. The current generation is Qwen3.8, from August 2026, and on November 16, 2026, Qwen4-Max, the first Qwen4 model, reached the API.

Is Qwen from Alibaba?

Yes. It’s built by Alibaba Cloud’s Qwen team. The official API is sold by QwenCloud, whose website is signed by Intelligent Cloud Computing (Singapore) Private Limited.

What is Qwen good for?

Agents and automations that generate a lot of text (because of the output price), running a model on your own server with Apache 2.0 weights, long documents of up to 1 million tokens via the API, image and video understanding, and fine-tuning small models on your own data.

Does Qwen cost money?

The chat is free and the open weights are free to download. The API is pay-as-you-go, in dollars: $0.15 per million input tokens and $0.47 output with Qwen3.8-Flash; $0.50 and $3 with Qwen3.8-27B; $2 and $6 with Qwen3.8-Max; $1.50 and $4.50 with Qwen4-Max (Singapore list prices). The Token Plan subscription starts at $6 a month.

Is Qwen AI free?

Largely, yes. Qwen Studio is free, the Apache 2.0 models (like Qwen3.8-27B) are free to download and use, and the API gives new accounts a trial quota of about 1 million tokens per model for 90 days, only in Singapore and only after you add payment details. After that, the API is pay-as-you-go.

Can you run Qwen for free?

Yes, if you run it yourself. The Apache 2.0 models are a free download from Hugging Face or ModelScope and run with vLLM, SGLang, llama.cpp or MLX; the only cost is your own hardware. The large 3.8 models also have downloadable weights, but at a size that calls for serious infrastructure and under custom licenses: internal use is allowed, but if your business is offering models as a service or AI work assistants, you may need a separate license.

How do I access Qwen Coder?

It depends which one you mean. Qwen Code is the official open-source terminal agent (qwen.ai/qwencode), and the Qwen3-Coder-Next model weights are on Hugging Face under Apache 2.0. Google also lists coder.qwen.ai under the name Qwen Coder; I haven’t verified what it offers or what it costs. If you already use Claude Code, Codex or Cursor, Token Plan lets you plug Qwen models into them.

Which is better, Qwen or Claude?

I don’t have an independent comparison I can cite, so I won’t claim one beats the other. They’re different bets: Claude is a closed model with its own ecosystem, while Qwen also gives you open weights you can run yourself. Test both on your actual task before you decide.

Is Qwen as good as ChatGPT?

Same answer: without a benchmark I can verify, I won’t claim it. What I can say is that Qwen gives you Apache 2.0 models you can download and run on your own server, and an API whose fast model costs $0.15 per million input tokens in Singapore.

Is Qwen or DeepSeek cheaper?

On the fast models, at Singapore list prices, they tie on input at $0.15 per million tokens, but Qwen3.8-Flash is cheaper on output ($0.47 versus DeepSeek’s $0.60 off-peak) and has no peak-hour rate, while DeepSeek doubles its price at peak hours.

Is using Qwen safe?

It depends on the region and scope you choose in QwenCloud. Singapore processes outside mainland China; in Frankfurt you can pick an EU-only scope, but the default workspace uses the Global scope, which can run on any node, mainland China included. I couldn’t verify the data retention policy for the pay-as-you-go API or the chat. For sensitive data, the safest route is running the open weights on your own infrastructure.

What’s the difference between Qwen3.8-27B, Flash and Max?

Qwen3.8-27B is a dense model with open weights under Apache 2.0. Qwen3.8-Flash is the API’s fast, cheap model (its experimental open version is Flash-Next). Qwen3.8-Max is the API’s large model, built on Qwen3.8-2.4T-A95B, whose weights are released under a custom license.

What happened to Qwen 4?

Since November 16, 2026, Qwen4-Max (qwen4-max) has been available through QwenCloud’s pay-as-you-go API: $1.50 per million input tokens and $4.50 output in Singapore, with a 1-million-token context window. It’s also on Token Plan. The Qwen4-Max-Preview, Token Plan only, will be retired when it ends on November 25, 2026. There are no open Qwen4 weights.

Can you help me build Qwen into my business?

Yes. I build open models and APIs like Qwen into agents and automations, and I train the team who’ll use them. You can see how I approach AI agents or message me directly.


Official sources checked on November 16, 2026: Qwen3.8 README · Qwen3 README (where chat.qwen.ai was still labeled “Qwen Chat”) · Qwen3.8-27B · Qwen3.8-2.4T-A95B and its license · Qwen3.8-Flash-Next · QwenCloud pricing and billing · Model pages for qwen3.8-flash, qwen3.8-27b and qwen3.8-max · Qwen4-Max and its explicit cache · Regions and scopes · Free quota · Token Plan and Token Plan Team · QwenCloud changelog · QwenCloud · Qwen Studio · Qwen Code. Alternative pricing: DeepSeek · Kimi · Mistral Large 4 · Mistral Small 4 · Meta.

[ Why StratoAI ]

StratoAI is a cross-industry AI implementation firm. An active portfolio of clients in real estate, specialist retail franchises, restaurant groups (AI work, not hospitality consulting), architecture, B2B distribution, private clinics and D2C ecommerce. A real production stack: Claude Code, n8n and agents running 24/7 — no demos, no slide decks.

A real case, anonymized under NDA: Port concession — diagnosis, business plan, brand identity, renders and website delivered in 15 days and submitted five days before the concession deadline. Executed with Focus Sprint™, the flagship methodology.

Want to bring Qwen into your operations?

If this tool fits your case, I’ll implement it inside one of StratoAI’s four lines: Signals, Automations, AI Agents or MicroSaaS. And if it doesn’t fit, I’ll tell you straight. A 15-minute Contact Call, no strings attached.

// Frontier stack

Models I evaluate · providers I run in production

Frontier + open-weight stack · continuous benchmarking · no vendor bias

Want to talk? I reply personally John Guerrero · StratoAI founder