Qwen3.8-Flash Announced as Open-Weight Qwen3.8 Flash

Alibaba Qwen (@Alibaba_Qwen) announced Qwen3.8-Flash in an Aug. 26, 2026, X post, describing it as a multimodal mixture-of-experts (MoE) model, an early preview of the Qwen4 architecture, and an open-weight release. The announcement also says that a production version will soon be available through the QwenCloud API. The supplied post is visibly truncated after the specification line “125B parameters + 51B N-gram…”, so the details below are limited to the information shown.

An open-weight model positioned as an early Qwen4 preview

The announcement gives Qwen3.8-Flash a role beyond that of a standalone model. It presents the release as an early look at the architecture behind Qwen4 while also identifying the model as multimodal and MoE-based. That makes the post relevant both to developers watching Qwen’s open-weight releases and to readers tracking the company’s next architecture generation.

The wording describes Qwen3.8-Flash as open-weight, but the announcement separately points to a future production version through QwenCloud. Based on the supplied text, the safe conclusion is that the model announcement and the planned hosted API are being presented together as two available or upcoming ways to engage with the release; the post does not provide further deployment details.

Planned QwenCloud API pricing

Alibaba Qwen says the production version will be available soon through the QwenCloud API at $0.16 per 1 million input tokens and $0.47 per 1 million output tokens. Those rates give developers an announced price reference for evaluating the hosted version when it becomes available.

The post also displays the line “125B parameters + 51B N-gram…”. Because the supplied announcement ends with an ellipsis, it does not explain the relationship between those figures or provide additional technical specifications. No further claims about the model’s performance, availability, or architecture can be established from the visible text.