Alibaba’s Qwen team opened the weights of Qwen3.8-Flash-Next on 2026-08-26, according to the news log in the official repository. The team is explicit about the purpose: this is a multimodal mixture-of-experts model that doubles as an early preview of the architecture that will underpin Qwen4. The stated precedent is Qwen3-Next, whose hybrid Gated DeltaNet plus Gated Attention design was released early for the community to examine and then carried through the Qwen3.5, 3.6, 3.7 and 3.8 series.
The architecture changes on four axes. Attention becomes a Gated DeltaNet plus Qwen Sparse Attention hybrid, where the sparse indexer selects important context at micro-block granularity rather than per token, cutting long-sequence attention cost. Gated Residual widens the residual stream into four branches with a dynamic gate controlling reads and writes. N-gram Embedding uses local context to look up a table that scales capacity for very little extra compute, and the table can be offloaded to host memory with asynchronous prefetching. Training uses the Muon optimizer with a refitted scaling law.
The size figures are unusual. The main model is 125 billion parameters, supplemented by a separate 51 billion parameter N-gram embedding table and a 4 billion parameter multi-token prediction component, with only about 6 billion parameters activated per token across 48 layers. Native context is 262,144 tokens, extensible to roughly 1,000,000. It is natively multimodal with a vision encoder, and runs in thinking mode by default with enable_thinking, preserve_thinking and reasoning_effort controls. Alibaba reports that training cost about one ninth of Qwen3.7-Plus while delivering superior capability on coding and office tasks.
Reported scores on the model card include 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 91.7 on GPQA Diamond, 84.5 on AndroidWorld and 95.7 on MathVision with chain-of-thought. The license is qwen-community-1.0 rather than Apache 2.0, so it is downloadable but not unconditionally permissive - a legal review is warranted before commercial deployment. Weights are on Hugging Face and ModelScope, and the model powers the new Standard mode on Alibaba’s QwenWork platform.