Re:Qwen4-Max
by seizethecheese on Sunday November 15, 2026 @09:48PM
(#50513416)
[delayed]
[ Parent
| on HN ]
Re:Qwen4-Max
by forgingahead on Sunday November 15, 2026 @10:02PM
(#50513520)
Qwen3.8 doesn't handle images?
Re:Qwen4-Max
by seizethecheese on Sunday November 15, 2026 @10:06PM
(#50513549)
Sorry that was confusing. Qwen 3.8 2.4T is faster than what we have on leaderboard but it doesn’t handle image input: https://openrouter.ai/qwen/qwen3.8-2.4t-a95b
In case it translates to Qwen 4 27B, I've been running Qwen 3.8 Flash Next (which while MoE (and bigger) is supposed to be a preview of Qwen 4's architecture) and it runs pretty solidly regardless of context window size and completes random deepswe challenges I throw at it locally (at Q2-Q3 quantization! so it fits on my shitty setup).
That is our experience as well at FP8 for 3.8 FN.
For many of our use cases it's currently far into the "good enough" range because external factors (such as process design) are now blockers.
Re:Qwen4-Max
by wronglebowski on Monday November 16, 2026 @08:41AM
(#50517506)
This, I continue to be amazed using Strata to serve IQ3_XXS. It's such a capable model for something that runs on (relatively) little hardware.
> but we couldn’t include it because it doesn’t handle images
Maybe I'm mixing up model variants, but Qwen 3.8 27B that I run locally totally does handle images. Surely the bigger one would as well?
Re:Qwen4-Max
by seizethecheese on Sunday November 15, 2026 @11:09PM
(#50513907)
[delayed]
Re:Qwen4-Max
by 0xbadcafebee on Sunday November 15, 2026 @11:38PM
(#50514103)
[delayed]
Is the speed due to a difference in the way it "reasons"? I was trying out a local version (3.8 27b) and watching the thinking, it's absurd how often it challenges itself. Simple facts get questioned multiple times, and it has to keep reminding itself that it already established some things as facts earlier.
Instructing it to ask any and all questions as necessary sends it into an existential crisis, because the question tool (oh-my-pi harness) suggested limiting the number of questions, and it couldn't figure out how to do both.
Aside from making our engineering prompt terse, I place a token budget on the thinking traces. Those traces are mostly meaningless(in the reasoning realm, not the usefulness realm), and just tell it to do something productive. For coding, this just makes it read doc, tests, etc.
I believe that's all reasoning levels do, but most inference engines should have thinking budget.
Re:Qwen4-Max
by seizethecheese on Monday November 16, 2026 @10:53AM
(#50518953)
[delayed]