Qwen3.8 Flash-Next activates 6 billion parameters from a 125-billion model

Alibaba’s Qwen team has released Qwen3.8-Flash-Next, a multimodal model that previews architectural ideas planned for Qwen4. Although the system contains 125 billion parameters, it activates only 6 billion for each token, reducing the computation needed for a response.

The mixture-of-experts model adds a 51-billion-parameter N-gram embedding layer, which stores common groups of words in a phrase-like dictionary. Qwen says that layer can remain in ordinary system memory rather than GPU memory at relatively low additional cost. Native context reaches 262,144 tokens and can be extended to one million with the YaRN technique.

Weights are available through Hugging Face and ModelScope, and the technical report is on GitHub. A production version called Qwen3.8-Flash is due through QwenCloud, priced at $0.16 per million input tokens and $0.47 per million output tokens; the API was not yet live when reported.

Alibaba claims Flash-Next beats larger rivals across several coding and office benchmarks while costing one-ninth as much to train as Qwen3.7-Plus. Those are vendor-reported benchmark results, and performance on a particular workload may differ substantially.