Qwen3.8-27B Has 6.76 Million Downloads Under Apache 2.0, but the 2.4T Flagship Carries a Different Licence
A.I. / news
Qwen3.8-27B Has 6.76 Million Downloads Under Apache 2.0, but the 2.4T Flagship Carries a Different Licence
Alibaba's August 14 release came as two open-weight models. The model card for the larger one names a custom licence, and at least one write-up of the release says both are Apache 2.0.
Qwen3.8-27B, Alibaba's open-weight multimodal model, had 16,972 likes and 6,758,884 downloads on Hugging Face when this story was filed on October 5, a count that puts it at the top of the site's trending list. Its model card lists an Apache 2.0 licence, which allows commercial use.
The model it was released alongside does not carry that licence. The card for Qwen3.8-2.4T-A95B lists qwen3.8-max, a custom licence, while The Decoder reported that both models ship under Apache 2.0.
The model cards are the primary source here, and they disagree with The Decoder's headline.
What the two cards say
Both were released on August 14 and both have downloadable weights. The 27B is a dense model, meaning every parameter is used for every token, and it accepts text, images and video. The flagship is a mixture of experts with 2.4 trillion parameters, 95 billion of them active per token, and it does not accept multimodal input.
| Qwen3.8-27B | Qwen3.8-2.4T-A95B | |
|---|---|---|
| Parameters | 27B dense | 2.4T total, 95B active |
| Licence on the card | Apache 2.0 | qwen3.8-max |
| Context window | 262,144 native, 1M with RoPE scaling | 262,144 native, 1,010,000 extended |
| Input | Text, image, video | Text |
The flagship uses 92 layers and 512 experts with 11 activated, per its card. Neither card states a hardware requirement. The 27B card points to vLLM, SGLang, TokenSpeed and Transformers for serving.
What the custom licence requires
The Decoder does not describe the Max licence, so its terms come from SQ Magazine, which quotes the licence text. Products serving more than 100 million monthly active users or $20 million in monthly revenue must "display the model name prominently in the user interface."
Companies running a Model as a Service or AI Work Assistant business with aggregate revenue above "$50 million across any consecutive 12 months" must "obtain a separate license from Qwen before Using the Software," per the same quotation. The Terminal did not retrieve the licence document itself.
A developer fine-tuning the 27B model is under Apache 2.0. A hosting company reselling the 2.4T model above that revenue line is under a negotiation.
Vendor-supplied scores
The benchmark figures on both cards are Qwen's own and no independent run is cited.
- Qwen3.8-27B61.7 score
- Qwen3.8-2.4T-A95B67.7 score
Source: Hugging Face model cards for Qwen3.8-27B and Qwen3.8-2.4T-A95B, accessed 2026-10-05
On GPQA Diamond the cards give 89.2 for the 27B and 92.6 for the flagship. The 27B card also lists 73.0 on Terminal Bench 2.1, 84.3 on OSWorld-Verified and 64.8 on WebArena-Verified. The card gives no statement about training data.
Why the licence split matters downstream
The 27B is already a base for other releases. Cloudflare's Clef is post-trained from Qwen3.8-27B and released under Apache 2.0, and Aleph Alpha's Kolibri also shipped under Apache 2.0.
The Decoder's headline, if it is wrong, would send a reader to the flagship believing the same rule applies. Alibaba has not said how it plans to enforce the revenue thresholds.
The Decoder lists a hosted version with a one million token context as forthcoming through Qwen Cloud, with no date given.
Sources
More in A.I.
- 01H Company's Holo4-27B Scores 61.7% on OSWorld 2.0, and Its Model Card Bars Commercial Use of an Apache 2.0 BaseThe September 28 release fine-tunes Alibaba's Qwen3.8-27B for computer use. The base allows commercial work; the derivative is CC BY-NC 4.0.
- 02ChatGPT Will Test Image Ads During Image Generation This Month, and Advertisers' Inventory Complaint StaysOpenAI added a visual ad format and more measurement partners on October 5. The supply problem a media buyer cites is one ad slot and about one-sixth of Google's daily volume.
- 03GPT-6.1 Sol Is Priced at One-Fifth of Astra, and Its System Card Rates Cyber CriticalOpenAI's Sept. 29 addendum also shows the model misrepresenting its own coding work more often than GPT-6 Astra did.
- 04Gemini 4 Argon Goes to Cyber Defenders First, With Broad Access UndatedGoogle priced its new frontier model at $2 and $10 per million tokens for an introductory period, then $4 and $20, and has not said when most developers get it.