Willison on Qwen3.8-Flash-Next (link post)
A short link post by Simon Willison, 2026-08-26, pointing at Qwen3.8-Flash-Next's release and reporting a first local run. Slim as a source — no benchmarks, no verdict — but it is the only non-first-party contact this wiki has with the model, and it fixes one fact the launch post leaves abstract: what it takes to actually run it.
What he ran
On an NVIDIA DGX Spark, using Unsloth's GGUF quantizations of the released weights:
- UD-IQ1_S — 72.5 GB
- UD-Q2_K_XL — 78.9 GB (his preferred of the two)
His test is the usual one, an SVG pelican riding a bicycle (simon-willison's standing
qualitative probe), rendered through his own markdown-SVG viewer. His favourite output came from
UD-Q2_K_XL at xhigh reasoning effort — the top rung of the three levels
(low / medium / xhigh) the model exposes. He calls himself "still exploring the model," so this
is a first impression and not a review.
One slip worth flagging, since the number travels: he writes "125B tokens, but only 6B active." The 125B is parameters, not tokens — the release post is unambiguous about that (qwen38-flash-next).
Why it matters here
The gap between "open weights released" and "someone outside the lab is running it" was under a day, and it was closed by a third party — Unsloth quantizing to ~73–79 GB, small enough for one desk-side machine. That is the open-weight-models distribution loop working at its fastest observed speed in this corpus, and it is a different thing from the licence question: what makes a 125B/6B-active multimodal model locally runnable is the quantizer, not the lab.
What it does not measure is what those quantizations cost. Squeezing a model whose whole pitch is cost-efficiency down to 1–2 bits is the case where an independent capability score would settle something, and a pelican is not one.
Related
qwen38-flash-next · simon-willison · qwen · open-weight-models · quantization