Qwen3.8-Flash-Next
- ID
- 18413
- Status
- summarized
- Published
- 27 Aug 2026, 7:52 AM
- Fetched
- 28 Aug 2026, 7:12 AM
- Provider
- Simon Willison
- Category
- developer-ai
- Original URL
- https://simonwillison.net/2026/Aug/26/qwen38-flash-next/
- Source URL
- https://simonwillison.net/atom/everything/
Summary
- Score
- 5.0
- Created
- 28 Aug 2026, 7:12 AM
- Tags
- Audience
- developersai-ml-learnersai-agent-users
What happened
Qwen released Qwen3.8-Flash-Next, a multimodal Mixture-of-Experts model with 125B total parameters but only 6B active, described as an early preview of the Qwen4 architecture. Simon Willison tested quantized versions (72.5GB UD-IQ1_S and 78.9GB UD-Q2_K_XL) on a DGX Spark via Unsloth, sharing early image generation results.
Why it matters
If you're running local models, the 6B active parameter count means this could run reasonably on high-end consumer hardware despite the 125B total size—but the quantized weights are still 72-79GB, so you need serious VRAM. The MoE architecture preview for Qwen4 is the signal worth watching if you're choosing model families for agent pipelines.
Discussion angle
Is a 125B/6B-active MoE model practical for local agent workflows in Malaysia where DGX-class hardware is scarce, or does this just push people toward API-based inference instead?