Alibaba’s Qwen team has released Qwen3.8-Flash-Next, an open-weight model built around a new architecture the company says will underpin its upcoming Qwen4 family. The mixture-of-experts model activates only 6 billion of its 125 billion parameters per token, plus a separate 51-billion-parameter lookup table the company calls N-gram Embedding.
A cheaper way to build bigger models
According to Alibaba, training Qwen3.8-Flash-Next cost only about one-ninth as much compute as its predecessor, Qwen3.7-Plus, while outperforming that model on coding and office-productivity tasks. The efficiency comes from four architectural changes: a hybrid attention mechanism combining Gated DeltaNet with Qwen Sparse Attention to compress long context cheaply; a “Gated Residual” design that widens the model’s internal information pathways; the N-gram Embedding table, which adds capacity with little extra computation; and a refined version of the Muon optimizer.
Alibaba’s own benchmarks put the model ahead of its earlier Qwen3.8-27B release and DeepSeek’s V4-Flash on several coding and agentic tests, and competitive with Anthropic’s Claude Opus 4.6 on some measures — claims that, like most vendor-reported benchmarks, haven’t been independently verified.
Following a familiar playbook
The release mirrors how Alibaba introduced Qwen3-Next ahead of the Qwen3.5 family: ship the new architecture as a smaller open-weight model first, let developers test it, then fold the lessons into the flagship release. Full weights and an FP8 version are already available on Hugging Face and ModelScope; Alibaba says a hosted “Qwen3.8-Flash” version will follow soon through its QwenCloud API, priced at roughly $0.16 per million input tokens.
The launch adds to a crowded month for Chinese open-weight models, with Qwen3.8-27B and the 2.4-trillion-parameter Qwen3.8-Max both shipping within weeks of it. The pace shows how quickly China’s AI labs are iterating in the open-weight space even though Qwen4 itself has yet to be announced.