Qwen3.8 Flash is Alibaba Cloud's flagship native vision-language model, presented as an early preview of the architecture expected in the next Qwen generation. It is a multimodal Mixture-of-Experts design pairing a 125-billion-parameter main network with an additional 51 billion parameters of offloadable N-gram embeddings, while activating only about 6 billion parameters per token. The model is described as open-weight and is aimed at high-volume applications, tool-driven workflows, and coding or co-working assistants that need strong reasoning over very long inputs and outputs.
The architecture was systematically reworked across four areas: attention, residual connections, embeddings, and optimization. The attention layer combines Gated DeltaNet, which compresses prior context history, with Qwen Sparse Attention, while residual paths run through a widened four-branch Gated Residual stream with dynamic cross-layer gates. Training uses the Muon optimizer together with a refitted scaling law, and Alibaba reports that this combination lowers training cost while improving coding and office-task performance compared with the prior generation. On Model Studio and Qwen Cloud the API is priced at roughly one RMB per million input tokens and three RMB per million output tokens, positioning the model as a price-performance option for long-context, agent-style deployments.
