We could really use Qwen3.8 in 27B, 35B, 122B and 397B sizes

Your post is getting popular and we just featured it on our Discord! Come check it out! You've also been given a special flair for your contribution. We appreciate your post! I am a bot and this action was performed automatically.

I just got off the phone with John Qwen and he says he'll get right on it

The brother of John Battlefield?

I’m here for this. Qwen3.6 27B at Q8 is sufficient for about half of what I do. The other half, I’ve been leaning on Qwen3.7 Plus or Max via API and cache control enabled and that gets me the rest of the way, helps me complete larger projects that require >262k, and doesn’t invoice shock me. I could frankenbuild out my machine to run 4x 3090 but why? A quantized 120B may not get me any closer than 3.6 27B at Q8, and 3.6 is so close. But a smarter 27 or 35 with extensible context? At the intelligence jumps we experienced this April to June ish? Yes, please.

This was my complaint in another thread -- seems like the strategy for open weight is make em so big no one can run them, and then we run them and rake in cash via api -- While they deserve to make a profit they should still honor the thing that made them popular, release capable small models for the little guy and small/medium sized business that want their own or on-site. At a certain point does it matter if its open if its cost prohibitive for any business or user to run them? Would love to see qwen 3.8 do another string of variant releases exactly as your title says. They were fantastic across the board. The retrains on those models like Nex N2 were also phenomenal.

We need a pitch that Alibaba will find persuasive

~120b needs some love rn

I would also love to see a release of new VL embedding models .

This has been discussed several times before. The chances of them releasing a small model are also small. The team (the main people) that was making those is gone and Alibaba leaderships said they are concentrating now on big monetizable solutions. They released the 3.6 27B and 35B ones just before they were canned, might even have been the last straw. I mean this is how it progressed: Qwen3.5 - 0.8B, 2B, 4B, 9B. 27B, 35B, 122B, 397B Qwen3.6 - 27B, 35B Qwen3.7 - nothing Qwen3.8 - ??? What do you think the chances are that 3.8 will have at least the same smaller releases as 3.6 or even more? Very slim. Releasing open weights models is one thing, but all of the latest big Chinese releases are not consumer or enthusiast friendly sizes. MiniMax, Kimi, GLM are all huge models where you need considerable hardware to run them. GLM is also similar to Qwen where there have been no small versions for the last 2-3 releases at all, nothing from the v5 for example.

What we really need is a good model that a typical consumer can afford to run on their local hardware. This is an expensive 'hobby' as it is. Anything over 30B requires hardware that is just unattainable\unaffordable at this point. I am happy for those that can afford paying a car worth of money for running LLMs locally. But I feel they are a minority even within this community. I prefer a model that "just works" and can be run on consumer level hardware than something only companies can afford. So yes, 27B please.

Meanwhile... Qwen: 27T, 35T, 122T, 397T??? 🤔 Comin' right up 😉

I agree. The Qwen 3.5 lineup was perfect. Model sizes to please basically everyone's needs

Sadly none of their current team members is saying anything about these model sizes which only means it's unlikely to happen. OP should keep using 3.6 until better models come along.

and 12-14b, 8-9b

Extreme hope for another 122b

40-50b a6b or something

Dont forget 4b, 7b, and the GOAT 9B!!!!

I don't get why there isn't a ~50 or 70b model. Going from 35 to 122 is such a massive leap for... What reason? And I think a 55B-A20B would be good. Or a 70B-A30B or something like that.

Instead of 2T+ models, continuing to release highly capable small to medium size LLMs would really help to keep this community vibrant. I would say "in addition to" rather than "instead of". Getting the full range from small to huge would be even better than only getting small models or only getting huge models. Why not both? The trend towards Chinese labs trying to match the Mythos class frontier with trillion parameter open weights models is not helping the local model community to innovate. Disagree. Sometimes those frontier-class open-weights (or sometimes even open-source) huge Chinese models are very beneficial both to other Chinese labs who then learn things from those huge models and/or distill down in size from them that helps them make better smaller models than they otherwise would've been able to make, or same thing but for some non-frontier American labs (i.e. could be helpful for labs like Poolside or Thinking Machines or Cohere or so on, who might then be able to make better small models than they otherwise would've been able to if China hadn't kept improving its 1T+ or 2T+ or however huge frontier models). Not caring at all about big models just because we can't run them on our local hardware is short-sighted, in my opinion. Lots of great new LLM tech can still come out of those are end up trickling down to smaller models that we end up benefiting enormously from later on. So I think a better attitude is to be happy about both, and say it would be great to also keep getting small and medium sized models from these labs, rather than to be like "screw big models, who cares about that stuff, they should stop working on those entirely and just only focus on small models". Not only is that extremely unrealistic, but it's not even necessary. If they did both, that would already be great (in fact, better than if they strictly focused on small models and nothing else, as it would likely improve the small models even faster if they did both than if they only did small models alone).

dont worry guys ill release a 32B model soon...

No Qwen 27B? Everybody go to Omar's post right now: https://x.com/osanseviero/status/2081398564345802934