Qwen3.8-Flash-Next
Blog · Watch/Read · (gathered 2026-08-29)
- Simon Willison is exploring Qwen3.8-Flash-Next, a new open-weights multimodal Mixture-of-Experts (MoE) model that serves as an “early preview of the architecture used in Qwen4.”
- The model is designed for efficiency, featuring 125B total tokens but only 6B active, which provides a “significant performance boost.”
- He is testing it on a DGX Spark using Unsloth quantized models, specifically trying the 72.5GB UD-IQ1_S (producing pelicans) and the 78.9GB UD-Q2_K_XL.
- His initial findings highlight its multimodal generation and reasoning capabilities, with a favorite example being an “xhigh reasoning effort one” from the UD-Q2_K_XL variant.