Key takeaways

  • Qwen3.8-27B is on Hugging Face under the Apache 2.0 license, which allows commercial use.
  • It is a dense 27B vision-language model with a 262,144-token native context, extensible to 1 million.
  • Other Qwen 3.8 models, such as Flash-Next, were not checked and may carry different terms.

Yes, Qwen 3.8 is open source in the sense most people mean: Alibaba's Qwen team published the weights of Qwen3.8-27B on Hugging Face under the Apache 2.0 license, which allows commercial use. The answer depends on which Qwen 3.8 model you mean, so this page covers the 27B model whose card we read on October 30, 2026.

Apache 2.0 is a standard permissive license. You can download the weights, run them on your own hardware, fine-tune them and build a product on them without asking permission or paying a fee, as long as you keep the license notice. That is looser than some "open-weight" releases, which add usage limits. Mistral's Large 4, for example, needs a separate commercial license if you run it as a hosted service (our explainer).

What the Qwen3.8-27B release includes

The Hugging Face page for Qwen3.8-27B lists the license as apache-2.0 and the format as Transformers weights in safetensors. The card describes the model as follows:

  • A dense model with 27 billion parameters. Dense means every parameter is used for every token, unlike the mixture-of-experts models that use only a slice.
  • A vision-language model. It reads images and video as well as text.
  • A native context of 262,144 tokens, extensible to 1,000,000 tokens.
  • Thinking mode on by default, with a reasoning_effort setting of xhigh, medium or low.

The card says the weights work with Hugging Face Transformers, vLLM, SGLang and TokenSpeed. As of October 30, the repository showed about 7.6 million downloads and more than 18,000 likes. Community builders have already published quantized GGUF versions, including one from Unsloth that carries the same Apache 2.0 license.

For setup steps, see our guide on how to run Qwen 3.8 locally.

What Alibaba claims about its performance

The model card includes Alibaba's own benchmark table. It lists Qwen3.8-27B at 73.0 on Terminal Bench 2.1 and 61.7 on SWE-bench Pro. In the same table, Qwen3.6-27B scores 63.4 and 53.5, and a model the card labels "Opus4.6 Max" scores 78.2 and 53.4. These are the vendor's numbers, run on the vendor's setup. Independent tests can differ.

Is it free to use?

The weights are free to download. Running them costs whatever your hardware or your inference provider charges. The card also says a hosted Qwen3.8-27B will come to Qwen Cloud, the official API service, with 1 million tokens of context by default. It describes that service as "coming soon", with no date or price.

What this does not cover

Qwen 3.8 is a family. The Qwen organization on Hugging Face also lists Qwen3.8-Flash-Next, and Alibaba has described a larger Max model. We did not read those cards, so do not assume they carry the same license. Check the license line on each model's own page before you build on it.

Qwen's own page for the 27B model is the place to watch for the hosted version and any license change.

Frequently asked questions

Is Qwen 3.8 free?

The Qwen3.8-27B weights are free to download. Running them costs your hardware or a provider's fees. The hosted Qwen Cloud version is listed as coming soon, with no price.

Is Qwen 3.8 multimodal?

Qwen3.8-27B is. The model card describes native support for image and video understanding along with text.

Can I use Qwen3.8-27B commercially?

Apache 2.0 allows commercial use. Keep the license notice, and check the card of any other Qwen 3.8 model before you rely on it.

Sources

  1. Qwen3.8-27B model cardHugging Face (Qwen)
  2. Qwen3.8-27B GGUFHugging Face (Unsloth)