Qwen3.8-Flash-Next
Alibaba · 180B parameters
Alibaba's open multimodal preview of the Qwen4 architecture, combining gated DeltaNet, sparse attention, mixture-of-experts layers and n-gram embeddings. It supports optional thinking and is distinct from hosted Qwen3.8-Flash; the card documents 262,144 native context tokens and a YaRN serving recipe extending to 1,000,000. Muon and AdamW are used for different weight categories.