Source · release
Qwen3.8-Flash-Next model card
Summary
The official model card for Qwen3.8-Flash-Next, an open-weight multimodal MoE model that the Qwen team calls an experimental preview of the Qwen4 architecture. It combines linear attention (Gated DeltaNet) with a new Qwen Sparse Attention (QSA). Unlike Qwen3.5, it is released under the custom "Qwen Community 1.0" licence rather than Apache 2.0. The card gives only "August 2026"; the day (2026-08-26) comes from news reports.
Key points
- 125B parameters, 6B active, plus 51B of n-gram embeddings and a 4B multi-token prediction module.
- 48 layers in a 12×4 hybrid layout of Gated DeltaNet and QSA blocks. 512 experts (10 routed, 1 shared) and gated residual streams with four branches.
- Context of 262,144 tokens natively, up to 1M with RoPE scaling.
- Accepts text, image, and video input.
- Self-reported scores: SWE-bench Pro 62.5%, GPQA Diamond 91.7%, AndroidWorld 84.5%.
- Runs on Transformers, vLLM, SGLang, TokenSpeed, and llama.cpp. Community projects run it on a single RTX 4090.
Why it matters
Shows the direction of Qwen4: sparse attention on top of hybrid linear attention, with few active parameters. The move away from Apache 2.0 needs watching.