Glonce

Qwen releases Qwen3.8-Flash-Next, an early preview of Qwen4

modelsThe Glonce editorsSimon Willison🔥76Glonce rating76Impact65Novelty78Relevance100Surprise55Players70Credibility85

Qwen has released Qwen3.8-Flash-Next, an open weights multimodal mixture-of-experts (MoE) model. A description quoted in Simon Willison's post about it, whose author is not named there, calls it "a multimodal MoE model that also serves as an early preview of the architecture used in Qwen4," Qwen's next major model generation.

The model is large: 125B in total, though the post's own label for that figure is "tokens" rather than the "parameters" more commonly used to size a model, with only 6B of that active at once. The post says that gap between total and active size is what gives the model a significant performance boost.

Willison has been running the model on a DGX Spark using quantized builds made by Unsloth: a 72.5GB UD-IQ1_S build, which produced output linked as "these pelicans," and a 78.9GB UD-Q2_K_XL build, with output linked only as "these." The post says the favorite result so far came from running the UD-Q2_K_XL build at a reasoning effort setting labeled "xhigh," without explaining what that setting changes.

The post is a short note framed as unfinished: "I'm still exploring the model." It gives no separate release date for Qwen3.8-Flash-Next itself, only the date of the post, 26 August 2026; no benchmark scores or comparisons to other models; and no license terms. Beyond crediting Unsloth for the quantized builds, it does not say where the model can be downloaded.

Key facts

  • Qwen has released Qwen3.8-Flash-Next, an open weights multimodal mixture-of-experts model that the post describes as an early preview of the architecture used in Qwen4.
  • The model totals 125B, a figure the post itself labels "tokens" rather than "parameters," with only 6B active at once.
  • The post says that gap between total and active size gives the model a significant performance boost.
  • Willison has run Unsloth-quantized builds of the model on a DGX Spark: a 72.5GB UD-IQ1_S build and a 78.9GB UD-Q2_K_XL build.
  • Willison's favorite output so far came from the UD-Q2_K_XL build run at a reasoning effort setting labeled "xhigh," which the post does not explain.

Why it matters

Qwen3.8-Flash-Next matters less as a finished product on its own than as a signal of what is coming: the description quoted in the post says it doubles as an early preview of the architecture behind Qwen4, Qwen's next major model generation. Its mixture-of-experts design, 125B in total with only 6B active at a time (the post's own label for that figure is "tokens," not "parameters"), is the kind of sparse activation the post credits with giving the model a significant performance boost. It also ships with open weights, which is why Unsloth was already able to produce the quantized builds Willison used to run it.

Who it affects

Developers and researchers tracking Qwen's roadmap get an early look at architectural choices that may carry into Qwen4. Anyone running open weight models locally, including on hardware such as a DGX Spark, is a direct audience for the quantized builds Unsloth has already produced. It also involves tool makers like Unsloth itself, whose quantization work is what let Willison run the model at all.

How to use it

Qwen3.8-Flash-Next is released with open weights. The post does not say where to download the model or its quantized versions, beyond naming Unsloth as the maker of the builds tried, and it gives no license terms or pricing. Willison ran two of Unsloth's quantized builds: a 72.5GB UD-IQ1_S build and a larger 78.9GB UD-Q2_K_XL build, the latter at a reasoning effort setting labeled "xhigh," on a DGX Spark.

How solid is it

Everything here rests on Simon Willison's own short post, framed as a work in progress: "I'm still exploring the model." It quotes a description of Qwen3.8-Flash-Next without saying who wrote it, though the "(via)" link in the title suggests it may trace back to Qwen's own release material; the post itself does not confirm that. The 125B and 6B figures come directly from the post, which labels them "tokens" rather than the "parameters" more commonly used to size a model. No benchmark scores, comparisons to other models, or license terms appear anywhere in the source.

Risks and caveats

The claim that the model previews Qwen4's architecture, and the claim that its sparse activation gives it "a significant performance boost," both come from the post itself rather than from independent testing or benchmarks. No release date is given for Qwen3.8-Flash-Next itself, only the date of this post, 26 August 2026. The post also does not explain what the "xhigh" reasoning effort setting controls or how it differs from other effort levels, and the example outputs it links to, "these pelicans," "these," and the xhigh result, are not described beyond those labels: they are linked images, not text discussed in the post.