Qwen AI explained: Alibaba’s models, Qwen Chat, Qwen Code and the image editor

Plans, prices and features checked on the official pages on November 1, 2026.

Qwen is the AI model family from Alibaba. It covers chat and reasoning models, image generation and editing, speech, and a terminal coding agent. Some models are open-weight and free to download, and others are sold only through Alibaba’s cloud.

The current generation is Qwen3.8. Its open-weight models include a 27B model under Apache 2.0, which is the practical choice for running locally. The top-end Qwen3.8-Max is a paid cloud model.

What Qwen is and where it comes from

Qwen is built by Alibaba and published on Hugging Face under the organization name Qwen. That page describes the group as one that releases large language models, large multimodal models and other AGI-related projects. So when people ask “who is Qwen AI” or “where is Qwen AI from”, the answer is Alibaba, a Chinese technology company, and the models are served from Alibaba Cloud.

The name covers many separate products. There are general language models, vision-language models (the VL line), omni models that take audio and video, text-to-speech, embeddings, image models and a coding agent. Choosing “a Qwen model” always means choosing one of these.

Which Qwen models exist and which one to pick

Alibaba Cloud’s Model Studio lists the models it serves. These are the main ones as of today:

• qwen3.8-max: the flagship, described for complex reasoning and coding and for visual understanding.
• qwen3.7-plus: balances performance, speed and cost, and also handles visual understanding.
• qwen3.8-flash: the fast, low-cost option.
• qwen3.8-omni-flash: audio and video analysis with text output, plus a realtime variant for live audio and video conversations.
• qwen-image-3.0-pro: image generation.
• qwen-audio-3.1-tts-plus, -flash and -next: text-to-speech. The Next variant is described as generating complete sound scenes with voices, sound effects and ambient sound.

The page does not list a separate Coder or VL model. In the current lineup, coding and vision sit inside the main Qwen3.8 and Qwen3.7 models. Older names such as Qwen2.5, Qwen2.5-VL and Qwen Coder are earlier generations, so check their model cards on Hugging Face before choosing one.

A simple way to choose:

• Best quality, cloud: qwen3.8-max.
• Cheap and quick, cloud: qwen3.8-flash.
• Best for coding, downloadable: Qwen3.8-27B. Alibaba describes it as improved for coding, professional work, research and long-horizon agentic tasks.
• Video generation: the Model Studio page lists no Qwen video-generation model, so no best one can be named. The Omni models analyze video but output text.

Are Qwen models open source, and where to download them

Many are open-weight, meaning you can download the weights, but the license differs per model. For the difference between open weights and fully open source, see our explainer on open-weight models.

The collections on Qwen’s Hugging Face page show what is downloadable now:

• Qwen3.8-27B: a 27B dense model with a vision encoder that takes images and video. The model card gives a native context of 262,144 tokens, extensible to 1,000,000, and an Apache-2.0 license.
• Qwen3.8-2.4T-A95B: a Mixture of Experts model with 2.4 trillion total parameters and 95B active per token. It is text-only and always runs in thinking mode. Its license is listed under its own name, “qwen3.8-max”, not Apache 2.0, so read it before commercial use.
• Qwen3.8-Flash-Next: an image-text-to-text model listed at 180B parameters, with an FP8 variant.
• Qwen-Image-2.1 and Qwen-Image-Edit-2511: the image models covered below.

The Hugging Face page does not show a Qwen3.7 collection, and Alibaba’s pages say nothing about Qwen3.7 weights being released. Treat Qwen3.7 as a cloud model until Alibaba says otherwise.

Security-minded teams should also read our report on Cohere’s claim that Qwen3.8-Flash-Next tried to upload a file to a Beijing storage bucket from inside a sandbox. Cohere says it could not tell the cause and that the request never left the container. Alibaba’s response was not in the post.

Running Qwen locally

Yes, the open-weight models run on your own hardware. The Qwen3.8-27B card lists these options:

• Hugging Face Transformers, vLLM, SGLang and TokenSpeed for serving, with OpenAI-compatible API examples.
• Docker Model Runner.
• Quantized versions that work with llama.cpp, Ollama and LM Studio.

Thinking mode is on by default and can be turned off per request. Reasoning depth is set with reasoning_effort, with xhigh as the default, plus medium and low. The card recommends YaRN rope scaling for contexts beyond 262K tokens and warns it can affect performance on shorter texts.

The 2.4T model is not a laptop model. Even with only 95B parameters active, the full weights have to be stored and served.

Qwen Image 2.1 and the image editor

Qwen’s image models are the most searched part of the family. Yes, Qwen can generate images, but the language models above are not the ones doing it. Image generation comes from the separate Qwen-Image line.

Qwen-Image-2.1 is an open model that does both text-to-image generation and image editing. According to its Hugging Face card:

• The visual generation component has about 7B parameters.
• It supports up to 10 reference images.
• Local edits can be marked with circles, painted annotations or separate masks.
• It can produce transparent (RGBA) images and extract subjects from photographs.
• Typography is described as improved, with text-rendering examples in the showcase.
• The license is the Qwen Research License Agreement, not Apache 2.0.

Qwen-Image-Edit-2511 is the earlier dedicated editor, at about 20B parameters under Apache 2.0. Alibaba says it improves on 2509 with less image drift, better character consistency, better multi-person consistency, built-in popular community LoRA effects such as lighting control and new viewpoints, and some industrial design abilities. The 2512 and 2511 numbers are release versions of the image line, not separate products.

For the hosted API, Model Studio lists qwen-image-3.0-pro for image generation. The pages read for this guide give no price for it.

Qwen Code, the command line coding agent

Qwen Code is an open-source coding agent that runs in the terminal. Its GitHub page describes it as Apache-2.0 licensed and originally based on Google’s Gemini CLI. It also has a desktop app, a web UI, editor plugins for VS Code, Zed and JetBrains, and SDKs for TypeScript, Python and Java. It supports subagents, MCP and plan mode.

Install it with npm (Node.js 22 or newer is required):

1. Run npm install -g @qwen-code/qwen-code@latest. Homebrew works too: brew install qwen-code.
2. Open your project folder and start a session with qwen.
3. Run /auth inside the session to set a provider and API key.
4. For scripts and CI, run it headless with qwen -p "your prompt".

The repository lists OpenAI, Anthropic, Gemini and Qwen APIs as supported, and local models through Ollama or vLLM. It does not mention a free tier, so how much it costs depends on the provider you connect. For how Qwen’s coding models compared in 2025, see our Kimi K2 vs. Qwen 3 vs. Claude Code piece.

API pricing and why Qwen can be free

Qwen is not free everywhere. The weights of open models cost nothing to download, which is the likely meaning behind “why Qwen is free”, but running them costs hardware, and Alibaba sells its cloud models by the token. No official page describing Qwen Chat’s plans or limits was available, so none are stated here.

Prices checked on November 1, 2026. The qwen3.7-plus prices are marked as a limited-time 20% off. Input prices exclude context cache pricing. The free quota is valid for 90 days from activation, model release or application approval, whichever is later, and applies only in Singapore. Other regions have their own prices. For a comparison across vendors, see our AI API pricing tracker.

How to start with Qwen

1. Decide whether you want a hosted model or local weights.
2. For hosted use, activate Alibaba Cloud Model Studio and pick qwen3.8-flash to start cheaply, or qwen3.8-max for the hardest tasks.
3. For local use, download Qwen3.8-27B from Hugging Face, or pull a quantized build through Ollama or LM Studio.
4. For images, use Qwen-Image-2.1 with Diffusers (it needs a recent transformers and torch>=2.4.0) and load QwenImage21Pipeline. Call enable_model_cpu_offload() to reduce memory use.
5. For coding, install Qwen Code and connect your provider with /auth.

Limits to know before you commit

• Licenses vary. Apache 2.0 covers Qwen3.8-27B, but the 2.4T model and Qwen-Image-2.1 use their own terms.
• Benchmark numbers on the model cards are vendor-reported. For Qwen3.8-27B, Qwen reports leading scores on SWE-bench Pro (61.7) and OSWorld-Verified (84.3), but its own table shows Claude Opus 4.6 Max ahead on Terminal Bench 2.1, GPQA Diamond and HLE. Some benchmarks are in-house.
• Cloud models run on Alibaba Cloud, so data handling and region matter for sensitive work.
• Version names move fast. Qwen3.7, 3.8 and 3.8-Flash-Next arrived within months, so check the model card date.

Alternatives worth comparing

Qwen is one of several open-weight families. Our related coverage includes MiniMax M3.1 Flash and Tencent’s WeVisOne-VL 4B for vision. For coding specifically, the Qwen Code card itself says it can run against OpenAI, Anthropic and Gemini APIs, so it is not locked to Qwen models.

Common questions