Qwen3.8-Flash-Next

Alibaba open-sourced Qwen3.8-Flash-Next on August 26, 2026, giving developers an early look at the Qwen4 architecture ahead of the full model family. It’s a multimodal mixture-of-experts model with a 125 billion parameter backbone but only 6 billion active per token, alongside a 51 billion parameter n-gram embedding table and a 4 billion parameter multi-token prediction module. Context runs natively to 262,144 tokens and extends to a million. Alibaba says it outperforms the older Qwen3.7-Plus, especially on coding and office tasks, while costing roughly one-ninth as much to train and about a twelfth as much to run. Both a standard and an FP8 checkpoint were released at launch.