On August 26, Alibaba released a new model: Qwen3.8-Flash-Next, along with an FP8 variant. The unusual thing isn’t the model itself but the timing. Qwen calls it an early preview of the architecture that the whole Qwen4 family will be built on. In other words, the team is publishing the design ahead of time so the community can get ready.
The specs: a multimodal mixture-of-experts model with 125 billion parameters, of which only about 6 billion are active per token. That’s exactly what makes it fast and cheap to run. Native context sits around 262,000 tokens and stretches to a million. The weights are on Hugging Face under an open community license; the production version will run later through Alibaba’s cloud API – at 16 cents per million input tokens and 47 cents per million output tokens.
The interesting part is in the attention. Until now Qwen used a mix of Gated DeltaNet and Gated Attention. What’s new is Qwen Sparse Attention, or QSA. Instead of picking individual tokens to process, QSA works at the level of small blocks. That noticeably cuts latency on long contexts – and long contexts are exactly where transformers usually get expensive.
Simon Willison already ran the model through his usual test and had it draw a pelican on a bicycle, on a DGX Spark with quantized versions. In benchmarks Qwen3.8-Flash-Next lands near the top, in some categories level with much pricier models like Fable 5 or GPT-5.6 Sol.
What I find notable is less any single number than the pattern. Chinese labs ship fast, in the open, and at short intervals. While the big US providers often test their best models internally for months, Qwen already has its next-generation architecture on the table before the main model even exists.
For us, that mostly means one thing: if you like building locally or with open weights, you get a fast, long-context, multimodal model to play with here – plus a fairly concrete hint of where Qwen4 is headed.
Sources: Qwen: Qwen3.8-Flash-Next · Simon Willison · TechNode