Will there be any more Qwen3.6 series models?

[-]

TokenRingAI@reddit

Maybe Qwen4-80B-A6B Qwen4-160B-A12B

Reply

[-]

cafedude@reddit (OP)

Hmm... Do you know something we don't?

Reply

[-]

TokenRingAI@reddit

https://preview.redd.it/34ar62i3xr0h1.png?width=1448&format=png&auto=webp&s=b040ff3ebfe86652a463525936a1405eeb5428d8

Reply

[-]

cafedude@reddit (OP)

They're going to fly in from Japan?

Reply

[-]

The fact that they are hesitant to publish small models is nothing short of a tragedy for GPU-poor people. With every major Qwen release, you could get small models ranging from <1b, 4b, 8b, etc. I suppose now we will have to make do with Qwen 3.5 and Gemma 4 small models for a while. 🫤

Reply

[-]

oldschooldaw@reddit

I am too gpu poor to properly run 3.6 27b so I’ve been putting 3.5 9b through paces I thought it wouldn’t be equipped for. With harnesses, hands, feet and eye tooling, it’s doing things I assumed I’d need a much bigger and smarter model for. I know I am def getting sucked up in the wave of newest newest newest must have newest but it’s quite amazing how powerful what’s already out is. I’m sure this isn’t news to you but it’s certainly calmed my need for newer and better when we got great already. Plus I’m resigned to the fact I straight up can’t afford enough compute to play with bigger anyway

Reply

[-]

Snoo_28140@reddit

That is great, and a very good realization. But I wonder can you not run the 35b?

Reply

[-]

oldschooldaw@reddit

i was trying 35b a3b and it was just too slow for me. i am running a single 3060 on that box, where as 9b was coming back in appropriate times

Reply

[-]

onephn@reddit

6 or 12gb? i got qwen 30b a3b running on a machine with 8gb vram with something like 30t/s, with ik\_llama, also what runtime are you using?

Reply

[-]

oldschooldaw@reddit

Llamacpp, 12gb vram but I want 64k context, it’s absolutely crawling

Reply

[-]

Great_Economist7685@reddit

use LM studio and quantize your KV cache and make sure it's unified disable both "trymmap()" and "keep model in memory" and do something like q8 quantize the K/V cache. IMO quality of code > speed.

Reply

[-]

voyager256@reddit

I thought MoE models were faster with higher context.

Reply

[-]

admajic@reddit

Just find a small enough quant to have it loaded like q2 should run fatter

Reply

[-]

oldschooldaw@reddit

At that low a quant though is there even any benefit to having a larger param count?

Reply

[-]

Monad_Maya@reddit

Try Gemma4 26B, might be a bit better. IQ4_XS from unsloth is pretty decent.

Reply

[-]

woswoissdenniii@reddit

How good are you with half a brain an all the confidence? Sure there are savants who could paint your moms arse from a mile away… but are they good at not beeing smelly?

Reply

[-]

Snoo_28140@reddit

I see. Thanks for sharing!

Reply

[-]

KickLassChewGum@reddit

> I suppose now we will have to make do with Qwen 3.5 and Gemma 4 small models for a while. 🫤 Gemma 4 is more than capable enough for 95% of use cases people tend to throw at AI

Reply

[-]

ICYPhoenix7@reddit

Waiter my steak is too juicy and my lobster too creamy

Reply

[-]

mc_nu1ll@reddit

except all the only steak you can afford is the TV Dinner one

Reply

[-]

marutthemighty@reddit

Why, what happened? Are Alibaba quitting the Qwen 3.6 series of LLMs? Or are they upgrading to bigger and better versions?

Reply

[-]

Perfect-Flounder7856@reddit

Moving more towards cloud based and api usage.

Reply

[-]

marutthemighty@reddit

Hmm. I see.

Reply

[-]

Synthetic451@reddit

Damn, just when it was getting real good.

Reply

[-]

Perfect-Flounder7856@reddit

Someone will fill the gap. We are so early in this.

Reply

[-]

tired514@reddit

I'm of the belief that in the future countries will always be vying to offer the most popular local models because it's a pretty significant "soft power." If Qwen is the undisputed champion for a few years, they have an opportunity to train and pre-prompt it to hold more positive views of China. If Gemma takes the crown, it'll no doubt have more positive views on the US. Could apply to the companies themselves too. I don't mean they'll be instructed to spread propaganda or anything, but geopolitically I suspect models *will* be seen as a form of influence. Perhaps in the future if you ask both Gemma and Qwen "which country in the world produces the leading AI software" they'll both offer slightly different takes favouring one country over the other. Or "which country seems to be the most technologically advanced" - it's a complex question with many facets, and baking positive messages into your country's most popular model exports will be pretty enticing.

Reply

[-]

woswoissdenniii@reddit

It’s not about what you say. It’s what about you don’t know.

Reply

[-]

tired514@reddit

Heh. Very good point.

Reply

[-]

Perfect-Flounder7856@reddit

So does that lean towards qwen coming in house cloud base and leaving the open weight scene?

Reply

[-]

tired514@reddit

I think the opposite. I suspect they'll release 3.6-122B (or some medium sized variant) at some point soon-ish so that they don't lose their spot as the #1 (or at least absolute top tier) local model in that size range. The existing models that are out there today aren't *really* the limitation keeping cloud AI afloat.. the problem is hardware. That problem will be solved in the coming years, so even if every single local model provider stopped publishing, it wouldn't save the cloud industry. For 95% of users what we have today is already sufficient... if they had the hardware to run it. Whatever limitations are there can be worked around with effective harnessing. So from their perspective I'm guessing it's: no more open weight models, cloud only (lose relevant in the local scene which *is* the future), or maintain open weights and have *something* of value as users shift away from the cloud. They could refocus on harness development, advanced local model licensing, custom training, consulting, and (of course) hosting for those who still need it. But if they let Gemma take over in the local space they'll lose a lot of future interest, I imagine.

Reply

[-]

Pleasant-Shallot-707@reddit

Gemma

Reply

[-]

GrungeWerX@reddit

Not *this* Gemma. Qwen 3.6 left some **BIG** shoes to fill.

Reply

[-]

Pleasant-Shallot-707@reddit

I’m not saying this Gemma, but Google is positioned to be totally cool with improving Gemma for the community IMO

Reply

[-]

ang3l12@reddit

Honestly NVIDIA is the only company that benefits from releasing an open weights model, as it will sell more hardware with nvidia’s stamp of approval, and that’s where NVIDIA makes their money. Google might be able to, but they tend to kill good things.

Reply

[-]

CodeDominator@reddit

Would not trust Google. Google graveyard dwarfs everything.

Reply

[-]

marutthemighty@reddit

Hmm. Noted.

Reply

[-]

Hot_Turnip_3309@reddit

I went back to 3.5 27B

Reply

[-]

SnooPaintings8639@reddit

I never actually worked with 3.5, but I have many issues with 3.6 for longer context task. What benefits of 3.5 over the newer version can you see?

Reply

[-]

Hot_Turnip_3309@reddit

no issues with longer context tasks is why i switched

Reply

[-]

HEAVYlight123@reddit

same, 3.6 seems marginally smarter if any and just overthinks way more

Reply

[-]

somerussianbear@reddit

HERESY! GET HIM!

Reply

[-]

yeah-ok@reddit

Absolutely, we got a monk who's gone rogue pagan on us here.

Reply

[-]

Ok-Measurement-1575@reddit

lol, why?

Reply

[-]

Routine_Plastic4311@reddit

My bet is they shifted focus to Qwen4 or incremental multimodal stuff. Coder variant would be nice but feels like a long shot now.

Reply

[-]

LegacyRemaster@reddit

I'm so sad...

Reply

[-]

FormalAd7367@reddit

i hope there’s something small that works great in my 3090

Reply

[-]

Afynith@reddit

Hey. you might want to read through this repo [https://github.com/noonghunna/club-3090](https://github.com/noonghunna/club-3090)

Reply

[-]

FormalAd7367@reddit

thanks so much! learnt many times already

Reply

[-]

NNN_Throwaway2@reddit

As I keep trying to point out, the 3.6 27b blog post implied there will be no more 3.6 model releases. Of course, nobody wants to believe it, so I’m not going to waste time arguing. We’re going on three weeks with nothing, if you still want to believe the qwen team is just still working on it, it isn’t my place to disavow you of that notion.

Reply

[-]

YRUTROLLINGURSELF@reddit

I shouted from the rooftops that that twitter poll was a measurement of consent, not actually a question of model preference, and only got downvoted. The humans left over here are not sending their best.

Reply

[-]

NNN_Throwaway2@reddit

Yeah, I got downvoted and was confronted by truculent people demanding I produce "proof" when I posted this several times in the past. Kind of funny the exact same info gets 100 net upvotes now.

Reply

[-]

IoannisHere@reddit

What blog post? This one? [https://qwen.ai/blog?id=qwen3.6-27b](https://qwen.ai/blog?id=qwen3.6-27b) Are you referring to this line? "With Qwen3.6-27B joining the roster, the Qwen3.6 open-source family now offers a **comprehensive range of models**, underscoring a generation where agentic coding achieved breakthroughs across every scale — from the 3B-active Qwen3.6-35B-A3B to the API-accessible Qwen3.6-Plus and Qwen3.6-Max-Preview. We are **grateful for the community’s feedback** and look forward to seeing what you build with these models. **Stay tuned for more from the Qwen team**!" It's ambiguous, but calling it a "comprehensive range of models" and thanking the community for their feedback makes it sound like that's it. The community has spoken, we delivered. And the "stay tuned for more" can be in any direction, closed or open. Maybe we have only us to blame for not getting a 122B-A10B...Mind you the 27B (and its MoE brother) have much broader applicability, it makes sense to release those as the "medium-sized versions".

Reply

[-]

Tagedieb@reddit

Also note how they specifically point out the coding capabilities, including in the title of the blog post. This shows that they don't consider a specific/separate coding model.

Reply

[-]

grumd@reddit

I think 3.6 was already trained very heavily towards agentic coding. Gemma4 is more of an all-rounder if you're looking for that.

Reply

[-]

IoannisHere@reddit

Right, I get the same reading from this. I see a recent trend with most AI labs not releasing specialists models. Thing is, this may reflect a training methodology breakthrough, where generalist models outperform fine-tunes. It kinda makes sense. For a model to be a good coder in the real world, it needs to have a solid understanding of diverse topics across domains of knowledge. With improved prompt adherence, we may no longer need coding-specific models. Or, all recent releases are really coding-agentic focused anyway, with mostly retained capabilities elsewhere.

Reply

[-]

Maximum@reddit

How about 3.7?

Reply

[-]

NNN_Throwaway2@reddit

There has been no announcement of future Qwen series models, let alone open weights.

Reply

[-]

mp3m4k3r@reddit

"3.6.1 wen?" /s

Reply

[-]

NNN_Throwaway2@reddit

Tru

Reply

[-]

FatheredPuma81@reddit

NGL I kind of feel like the Qwen3.6 models we did get were only released because they already invested so much into making them... I hope Qwen 3.7/4 has open source releases but it's looking bleak. Way off topic but I'm curious what route open source models are going to ultimately go to be viable by the companies making them. I thought Qwen's strategy was to release good quality small models to basically advertise their larger cloud models but I guess I was wrong or that has changed.

Reply

[-]

robberviet@reddit

Unlikely that everything is out now. The best we could hope for is something like Qwen-next at 72-80B MoE range.

Reply

[-]

Dangerous_Fix_5526@reddit

21B Qwen 3.6 is here: [https://huggingface.co/DavidAU/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking](https://huggingface.co/DavidAU/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking) Working on 14Bish / 10B ish currently.

Reply

[-]

crantob@reddit

Maybe it's not easy for them to improve on 3.5/3.6 27b You haven't begun to really leverage it yet, have you?

Reply

[-]

mivog49274@reddit

Qwen will continue to deliver. Frontier ain't their hood. If they leave the local/mid-sized model space, they will just disappear. Deepseek V4 "preview" has now vision capabilities and is blowing their 3.6-Plus and 3.6-Max out of the water. Qwen team is known to be composed of kings of optimization and power-per-parameter ratio. Either they break through the frontier but I think there are a few chances, either they continue on the bet on the decline of the cost of intelligence, as said by Kilpatrick, and keep their strategic edge on the field.

Reply

[-]

Technical-Earth-3254@reddit

I'm still hoping for the 3.6 397B open weights

Reply

[-]

Puzzleheaded_Base302@reddit

if you can run qwen 397B locally, you can also do deepseek-v4-flash, then why bother with qwen.

Reply

[-]

Long_comment_san@reddit

I'm pretty sure open coder models will start to diminish as they directly compete with cloud AI revenue. It's time, people, people actually start to care about profitability and qwen internal movements were pretty much expected by me and I bet they're not the last. Not sure about other models like 3.6 122b though. I'm pretty sure it's important as lite version of larger model (you won't use 35b for this I think lol) but I'm a little unsure here. Personally I think we're gonna hit a plateu for a while in "sane" category. We have a bunch of MOE 120b models - mistral, nemotron, qwen. It makes veeeeery little sense to bother with this range further, it's quite saturated. Everyone must be doing 300-700b range by now and next 3 months will tell us whether I'm right. I'd like to see 9b model. Each +10% in that 8-12b range is actually enormous as a lot of people run it locally. Qwen 9b finetune numbers on HF are CRAZY, it's in the thousands last time I checked. Well, I guess 35b took 9b place this round from 9b 3.6. Personally I think 8-9b are just a tiny bit too dumb relative to 12b which don't take that much more resources to run but are substantially more intelligent. Mistral actually nailed 14b size (13.5+0.5 if I'm not wrong?), it's a perfect size for home use.

Reply

[-]

fantasticsid@reddit

The 9b runs very well on an 8GB GPU with room for a ton of context due to the recurrent layers, and is actually pretty smart. It just lacks world knowledge, but that's fine since it's really good at calling tools. Give it a web browser and it's amazing.

Reply

[-]

Alternative_Ad4267@reddit

I was thinking just that. You see, Qwen 3.6 27B at BF16 doesn’t require that much memory (relatively speaking), but it requires a decent amount of GPU power (one or two powerful professional cards). The medium models that are missing, due to Qwen’s 3.6 demonstrated capabilities are a threat even for their own makers, a Medium sized model that would require around 10k-15k USD investment to run decently. Right now, you can spend like 7k to 10k to run Qwen 3.6 27B at Q8 or BF16 with a decent speed. And then you will have to spend like 40k to run Kimi or DeepSeek models. There’s no in between. And it won’t be.

Reply

[-]

fantasticsid@reddit

I just spent $500 on MI50s to run it at FP8 in VLLM. Spending 7k seems... excessive.

Reply

[-]

sheetis@reddit

Curious what you consider "decent speed" at this point? With MTP and DFlash being available, there are some options to really get some speed out of the dense model. I think for well under 7k, you could run Qwen 3.6-27B at amazing speed: I run a couple of AMD 7900 XTX I got for \~$800 each for Qwen 3.6-27B * Q8\_0 Quant * 131072 Context * Tensor Parallelism * MTP Support on a test prompt script that is around 72k context prefill and around 3000 tokens generated, I get something like this: * \~680 tok/sec prefill * \~60 tok/sec generation The prefill is on the slower side, but with good prompt caching, it's barely noticeable. In real-world turns I often see stuff more like this: slot print_timing: id 0 | task 10548 | prompt eval time = 19055.34 ms / 16808 tokens ( 1.13 ms per token, 882.06 tokens per second) eval time = 2578.00 ms / 192 tokens ( 13.43 ms per token, 74.48 tokens per second) total time = 21633.34 ms / 17000 tokens draft acceptance rate = 0.86792 ( 138 accepted / 159 generated) slot print_timing: id 0 | task 10547 | prompt eval time = 3177.81 ms / 2381 tokens ( 1.33 ms per token, 749.26 tokens per second) eval time = 44588.20 ms / 3143 tokens ( 14.19 ms per token, 70.49 tokens per second) total time = 47766.00 ms / 5524 tokens draft acceptance rate = 0.74417 ( 2170 accepted / 2916 generated)

Reply

[-]

xienze@reddit

> Curious what you consider "decent speed" at this point? I run it at BF16 on my RTX Pro 6000 and it's something like 40ish tokens/second, which is plenty for agentic use. The main problem it has is doing an insane, absolutely insane amount of thinking, even for simple questions (like give me a bash script to rename files matching this pattern and such).

Reply

[-]

yeah-ok@reddit

Vulkan I take it?

Reply

[-]

sheetis@reddit

These are ROCm. Tensor parallelism doesn't really work in Vulkan and the performance in layer and row parallelism wasn't keeping up for me. This is TP + MTP PR on 2 GPU (7900 xtx). I think tensor mode layer could get close, but so far these numbers are the best/most reliable numbers.

Reply

[-]

Alternative_Ad4267@reddit

Yeah, my minimum decent generation speed would be 60 tokens per second at Q8. BF16 if you really need that 1%-2% theoretical extra precision. That speed is the sweet spot for a coding tool that feels just like the cloud ones. Q4 is widely publicized here but in my tests is not that good. Q6 is decent enough if you don’t have enough hardware, maybe even a shot to Q5. I don’t like KV cache quant, that degrades accuracy in my tests. MTP is good at 2 or 3 tokens (decent speed up and maintain accuracy). Flash Attention fine and even required for 262k tokens context.

Reply

[-]

sheetis@reddit

I really wish I could fit BF16 myself, but for me I've found Q8\_0 is the bare minimum I can accept. I have a 3rd 7900 XTX in my gaming system which might end up getting repurposed for the LLM workstation, and gaming can downgrade to a an RTX 2070 I have sitting around.

Reply

[-]

fantasticsid@reddit

They seem to be dropping something once a week. - WebWorld (based on qwen 3, but novel) - That Sparse AutoEncoder thing - The 27B - The A3B I'm hoping we see the 9B at some point, because the 3.5 9B hits a particular performance/functionality spot on low-end GPUs.

Reply

[-]

Organic_Scarcity_495@reddit

they're probably cooking the 122B but waiting on the training run to finish. the gap between 35B and 235B is too wide for their lineup — a 122B coder variant would slot perfectly.

Reply

[-]

ttkciar@reddit

I'd like to see Qwen3.6-9B for dataset augmentation tasks, but Qwen3.5-9B is fine in the meantime. In my more optimistic moments, I hope that they are trying to fix the overthinking problems before releasing Qwen3.6-122B-A10B, but that seems increasingly unlikely. More likely is that the Qwen team is re-learning how to work as a team, since they recently shuffled out/in a few members, and we won't see anything for a while.

Reply

[-]

alex_bit_@reddit

Hoping for a 3.6 122b.

Reply

[-]

Faktafabriken@reddit

Dense!

Reply

[-]

ttkciar@reddit

"We have dense at home!" (Dense at home: Qwen3.6-27B)

Reply

[-]

q5sys@reddit

I was too, but it looks like we're not going to get one.

Reply

[-]

blastbottles@reddit

I would like to see more development on Qwen 4 and what that might be

Reply

[-]

GrungeWerX@reddit

At this point, even if it's marginally better...it's a major win for us.

Reply

[-]

inthesearchof@reddit

All that are wondering should be asking at [https://x.com/Alibaba\_Qwen](https://x.com/Alibaba_Qwen) & [https://x.com/ChujieZheng](https://x.com/ChujieZheng) . Let them know.

Reply

[-]

a_beautiful_rhind@reddit

They had a major restructure so it doesn't look good.

Reply

[-]

cafedude@reddit (OP)

Sure, but the 3.6 models we have were released after that restructure.

Reply

[-]

a_beautiful_rhind@reddit

Yea but they were cagey about even giving the 397b.

Reply

[-]

bonobomaster@reddit

Maybe they started training before the restructuring and were so far into the process, that management greenlit finalizing and releasing the model. Dunno, just talking out of my ass.

Reply

[-]

my_name_isnt_clever@reddit

So is every single person in this thread. At least you're honest about it.

Reply

[-]

UncleOxidant@reddit

It's possible that the 3.6 models we have were already in-flight prior to the restructure. I guess we'd have to go back to the poll they did on X about which 3.6 model to release first to see how that date coincides with the restructure.

Reply

[-]

Ariquitaun@reddit

What was the result of the poll?

Reply

[-]

Last_Mastod0n@reddit

People wanted to see the dense 27b models before any mixture of expert models

Reply

[-]

misha1350@reddit

No

Reply

[-]

Prof_ChaosGeography@reddit

They released the 36b 3ab Moe and it showed a great small moe model, and did great in benchmarks. Then the 27b was released and it only did slightly better then the previous version. It's likely the other models didn't offer a decent improvement in benchmarks so they decided to hold the release rather then face mean people on the Internet

Reply

[-]

Vicar_of_Wibbly@reddit

Crossing my fingers for 397B over here.

Reply

[-]

Regular-Forever5876@reddit

Hinestly the n'est sise overall is around 80B dense. Thoughts anyone? 🙂

Reply

[-]

IoannisHere@reddit

Their poll included the 122B-A10B: [https://x.com/ChujieZheng/status/2039909917323383036](https://x.com/ChujieZheng/status/2039909917323383036) People voted for the dense 27B. Their wording "We are planning to open-source the Qwen3.6 models (particularly medium-sized versions)", gives me hope that at least the 122B-A10B will be released.

Reply

[-]

tamerlanOne@reddit

La famiglia 3.6 è molto performate e copre un ampia fascia di utenti ed hardware diversi. Fare uscire altri modelli significa sovrapporli tra loro senza un reale beneficio tangibile creando confusione. Meglio che le energie di sviluppo vadano alla nuova famiglia qwen 3.x o 4 😉

Reply to Post

102 Comments