Qwen Labs has released Qwen3.8-Flash-Next, an open weights multimodal MoE model serving as a preview for the future Qwen4 architecture. The system contains 125 billion parameters but activates only 6 billion during inference, allowing it to run on consumer hardware while maintaining high performance. Simon Willison tested the model on an Nvidia DGX Spark using Unsloth quantised versions, specifically the 72.5GB UD-IQ1_S and 78.9GB UD-Q2_K_XL variants. His initial outputs included detailed illustrations of a pelican riding a bicycle with a fish in a basket.
This release demonstrates how mixture-of-experts designs enable larger models to operate efficiently on smaller devices without sacrificing quality. The active parameter count remains low enough for local deployment while the total size suggests future scaling potential. The model also incorporates xhigh reasoning capabilities that improve complex task handling.
* Total parameter count is 125 billion
* Active parameters during inference are 6 billion
* Tested on Nvidia DGX Spark hardware
