← Back to the A-Z glossary
Model · Alibaba · 2026

Qwen 3

Qwen 3 is Alibaba's flagship model, released in early 2026. It's open-weight (you can freely download its weights) and trained for 118 languages, so it's a real alternative when you want to step away from the American models and DeepSeek, especially for niche languages.

Model 4 min read Updated 2026-05-25
— What it is

Qwen 3, in plain words

In plain words: it's Alibaba Cloud's family of models, and you can run it on your own servers thanks to its open license, unlike Llama.

Qwen 3, released in early 2026, is their current generation. Several sizes are available (from 0.5B to 235B parameters), it's multimodal on the Pro versions (text, image, video), and it was specifically trained for 118 languages — solid performance in French, German, Arabic, and Spanish.

It's under the Apache 2.0 license on most versions, which makes it a truly open model, unlike Llama.

What I use it for

  • Niche multilingual: if I need to handle content in Arabic, Thai, or Vietnamese, Qwen is often better than the American models.
  • High-volume pipelines: gentle pricing, solid performance, ideal for bulk processing.
  • Multilingual code: comments in French, docs in English, Qwen juggles it well.
  • Self-hosting: with open weights, I can run it on my own servers without sending my data to Alibaba.
  • Diversifying your stack: not depending only on Claude/GPT, keeping a plan B.

How it compares

vs DeepSeek V3.2: similar pricing, two Chinese open-weight models. DeepSeek is better at code and reasoning, Qwen at multilingual (118 languages vs ~50 for DeepSeek).

vs Mistral Large 3: similar pricing, two genuinely open-source models (Apache 2.0). Mistral is better at professional French, Qwen is better at Asian and Arabic languages.

vs Llama 4: Qwen has a more open license (Apache 2.0 vs Meta's restricted one). Llama has a more mature ecosystem when it comes to tools.

What it costs

API pricing as of May 25, 2026 (source: alibabacloud.com):

  • Input: $0.30 / million tokens
  • Output: $1.80 / million tokens
  • Model weights: free on Hugging Face, Apache 2.0 license
  • Via Together AI or Fireworks: similar pricing, a Western host if you want to avoid Alibaba Cloud

My take

I don't use it day to day — Claude and Gemini cover my needs. But it's a model I test from time to time to stay up to date on what's happening in the open-weight world.

What I like: the Apache 2.0 license, the impressive multilingual range, and the fact that it's not DeepSeek (diversifying the Chinese sources).

What bugs me: the docs are in shaky English, there's basically no French community around it, and the consumer app isn't easy to access from Europe.

Quick questions

Is Qwen Chinese?

Qwen 3 is Alibaba's open-weight model released in early 2026, multilingual (118 languages) and under the Apache 2.0 license on most versions. Strong on niche languages, it can be self-hosted to keep your data private.

Qwen vs DeepSeek, which one should I pick?

DeepSeek for code/reasoning, Qwen for niche multilingual.

Is there a consumer app?

Yes, "Qwen Chat," but access from Europe is spotty. It's simpler through Hugging Face Spaces.

Checked on 2026-05-25 · next review 2026-11-25

Pricing and specs verified on Alibaba Cloud Model Studio. Personal use: occasional tests, not in production.

Want to keep going in order?

Head back to the full glossary, search a word, then open only the pages that deserve more than a short definition.

Open the AI glossary