One Qwen 3.8 name covers four models

On September 28, OpenRouter published a note asking what Qwen 3.8 is. It says Qwen 3.8 is Alibaba's generation from August 2026, and that four models share the name. The model you can download is not the model you can call.

You can download Qwen3.8 27B under Apache 2.0, and Qwen3.8 2.4T A95B under the Qwen3.8-Max License. Qwen3.8 Flash has no repository of its own. It is based on Qwen3.8-Flash-Next, which is downloadable under the Qwen Community License 1.0. Qwen3.8 Max has no public weights.

The release notes record the 2.4T-A95B weights arriving on Hugging Face on August 12, 2026, and the 27B weights on August 14. The September 28 note is separating these four artifacts. It is not announcing a model that launched that day.

Alibaba's model card describes Qwen3.8-Max as the hosted version of Qwen3.8-2.4T-A95B, with vision input, a non-thinking mode, a default context of 1,000,000 tokens, and built-in tools. Flash has the same relationship. Qwen3.8-Flash is the hosted version of Qwen3.8-Flash-Next, again with a 1,000,000-token default context and built-in tools.

The model ID qwen/qwen3.8-max is an alias. It currently resolves to qwen/qwen3.8-max-0902, the snapshot dated September 2, 2026. The note says to use the dated ID if a request should keep hitting that snapshot.

Prices are US dollars per million tokens, prompt then completion, from the OpenRouter catalog on September 11, 2026. Max is $2 / $6. Six of the seven 2.4T providers charge $2 / $6, and Venice charges $2.50 / $7.50. Fourteen providers serve the 27B, from $0.15 to $0.45 on input and $2.00 to $3.20 on output. Flash is $0.15 / $0.47.

Open weights have a native context of 262,144. In the catalog, the largest hosted window for the 2.4T is 1,048,576. Max, the 27B, and Flash list a largest hosted window of 1,000,000. Max's native context is not published. The 27B model card describes extending that model to 1,000,000 tokens with YaRN, and notes that static YaRN can hurt shorter inputs. Whether a provider extends the window, and how, depends on where the request is routed.

The 2.4T is an instruction-tuned model with 2.4 trillion total parameters and 95 billion active per token. It does not use Apache 2.0. The license grants use, copying, modification, distribution, sublicensing, sale, deployment, hosting, fine-tuning, and derivative works, subject to two conditions. If a commercial product or service built on the model has more than 100,000,000 monthly active users, or more than $20,000,000 in monthly revenue, the model name must be shown prominently in the interface. If you or your affiliates run a Model as a Service or AI Work Assistant business, and aggregate revenue exceeds $50,000,000 in any consecutive twelve months, you need a separate license from Qwen before commercial use. Internal use that does not expose the model, its outputs, or its capabilities to a third party is outside that condition.

Flash-Next is described as an experimental preview of the architecture that will underpin Qwen4. Its attribution condition matches the Qwen3.8-Max License. Any Model as a Service or AI Work Assistant business needs a separate license before commercial use, and that condition has no revenue threshold. Two of the three open-weight models therefore ship under licenses Qwen wrote that are not on the OSI approved list. Open weights means the checkpoint can be downloaded and inspected. It does not mean the license is an open-source license.

Qwen 3 is an earlier generation from April 2025. Qwen 3.5, 3.6, and 3.7 shipped between that release and Qwen 3.8. A license term from Qwen 3 does not describe these four models. The 27B is a dense model of 27 billion parameters, about 54 GB at bf16. Only 95 billion parameters of the 2.4T are active on a token, but all 2.4 trillion must be loaded. The repository README shows serving commands for the 27B only. The open weights of the 2.4T read text alone. On September 11, 2026, an image sent to qwen/qwen3.8-2.4t-a95b failed with HTTP 404 and the message that no endpoint supported image input. The same request to the 27B and to Flash succeeded. This piece did not recheck the prices or open the leaderboard.