Comment by ericd
5 hours ago
Thanks! That's very helpful as a way to discuss.
First off, I'd include Qwen flash-next and GLM 5.3 to show some of the other strong open weight models, and they predictably dominate it, but they're much larger. But, it shows up right next to DSv4 Flash 0731 on the overall index, and that's much larger. It's a great model! But then scroll down and hit Time Per Task, and you'll see that DSv4 Flash takes 3.6 seconds per task to Qwen's 21.1. That's what I meant when I said this:
>speed due to excessive thinking maybe to make up for the smaller amount of world knowledge baked in (qwen 27b's main issue iirc), etc - they're tuned for different things.
It can make up for its shortcomings by iterating a lot longer, and using way more thinking tokens. And that's a great trade if you don't have the vram to run the bigger models, but speed is pretty important for getting things done... And that's why DSv4Flash is great, too, despite being much larger, and scoring similarly on the intelligence index.
> Qwen flash-next and GLM 5.3 to show some of the other strong open weight models, and they predictably dominate it
Absolutely - no argument from me here. Bigger is very clearly a lever you can pull to get more out of a model.
> But then scroll down and hit Time Per Task, and you'll see that DSv4 Flash takes 3.6 seconds per task to Qwen's 21.1
Fair point, qwen definitely is slower - it’s a dense model, 27b params, vs a sparse 13b active params model - but the data doesn’t quite agree with what you’re saying about reasoning. I.e.:
> It can make up for its shortcomings by iterating a lot longer, and using way more thinking tokens
If you look at the total tokens generated, deepseek thought for 45k tokens and qwen thought for 48k. Barely a difference. The wall clock difference is all down to the speed of token generation, not the amount of reasoning done. At least when we are comparing deepseek and qwen 27b. The comparison swings more towards your position when it comes to the other models on the chart that reason for much fewer tokens.
So perhaps a hypothetical Qwen-27b-a13b could never rival deepseek’s larger model and the tradeoff is one of speed vs overall size - i.e. a small model needs more active params to compete than a big one does.
One data point that seems relevant to me is that the previous gen qwen Qwen3.6-27b was not so different in performance from its sibling model Qwen3.6-35b-a3b. We never got a qwen3.8-35b-a3b, but if we had, would the gap have stayed the same or gotten bigger? I.e. would the quality gains by improving training coming up against a hard limitation with 35b, or not.
Ah good catch on the total tokens, was going off vague memory there, and I thought people had gotten qwen 3.8 27b up to similar decode speeds as ds v4 flash.
>One data point that seems relevant to me is that the previous gen qwen Qwen3.6-27b was not so different in performance from its sibling model Qwen3.6-35b-a3b. We never got a qwen3.8-35b-a3b, but if we had, would the gap have stayed the same or gotten bigger? I.e. would the quality gains by improving training coming up against a hard limitation with 35b, or not.
Yeah good question, kind of shocking that a 3b active model would perform as well as a 27b dense.