Intel / article

Qwen3.8-Flash-Next

Source: simonwillison.net
Date: Published Aug 26, 2026

Context

Simon Willison reports trying two versions of Qwen's Qwen3.8-Flash-Next on a DGX Spark. Willison quotes the model as a release that also previews the architecture planned for Qwen4. The quantized versions store the model at reduced numerical precision. The August 26 post identifies a 72.5 GB build and a 78.9 GB build, concrete examples of the file sizes involved in a local run.

The report is an exploratory trial on that machine. Willison's favorite among the generated pelican drawings is a personal assessment, and the post provides neither measured generation speed nor a systematic comparison of the two versions' quality.

Terms in this piece · Glossary

quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.

Read the source simonwillison.net