Waiting on Qwen to drop those 3.7 models be like:

Posted by Porespellar@reddit | LocalLLaMA | View on Reddit | 28 comments

Mods please be kind. This was not “low effort”. It took me several minutes to find just the right waiting room gif to capture the sentiment of all us folks patiently waiting for our brothers and sisters in the east to hopefully drop some amazing new models on us.
I’m hoping for the 27b and 122b models, but I’ll be happy with whatever at this point. We need to see our little Capybara friend make an appearance here soon.

[-]

oldschooldaw@reddit

I haven’t even properly spun up 3.6 yet, I only last week or so truely gave 3.5 a go, this moves so fast..

[-]

Top_Break1374@reddit

Qwen has basically stopped oss since justin and others left the team. Qwen is doomed if you ask me.

[-]

sleepingsysadmin@reddit

im betting it'll be 4b, 9b, and 122b

[-]

snapo84@reddit

if the 9B can beat the 3.6 27B ... that would be OMG moment awesome... but i highly doubt that

[-]

sleepingsysadmin@reddit

i doubt it. it'll probably be in the qwen3.5 35b area probably.

[-]

u23043@reddit

It hasnt even been 30 days since 3.6 27B came out

[-]

Budget-Toe-5743@reddit

Hey, does anybody know how much memory each model will need?

[-]

BeautyxArt@reddit

3.6 is it that bad bro ..?

[-]

Juulk9087@reddit

Using 27B BF16 weights and BF 16 cache I can hardly get anything done. It needs an extensive amount of conversational training. Half the time it has no idea what you're talking about and it has no idea to infer or predict what you're talking about. So yeah it needs a lot of work. I'm hoping 3.7 trained on predictive and conversational data sets.

[-]

BeautyxArt@reddit

yes and most of times it give response way off about what you are talking about , with that wrong behavior ..it trust it self like it understand you , but it just output a mess of mess.

anyway 27b model is way way slower than the 9b one , bad idea i tried to move to the 27b while i have not good HW. but is it really 3.6 slower than 3.5 ? or behave differently (the code) ? same quant same code same setup , 3.6 ..is there any something (processing wise) different over 3.5 ?

[-]

Porespellar@reddit (OP)

I mean…. It’s not…. but 3.7 is a greater number than 3.6 sooooo 😂

[-]

Subject_Mix_8339@reddit

ayo we need a fucking MoE which splits nicely that fills most vram and ram. Pls have more active params 👏

[-]

nick_frosst@reddit

in the meantime, try this 😄 https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16

[-]

MoistRecognition69@reddit

trying to build a pipeline on the latest qwen model is exhausting

[-]

nunodonato@reddit

I'm hoping for ANYTHING. Still not sure if they will release open weights at all

[-]

challis88ocarina@reddit

given what's going on elsewhere, theres's nothing more coming for a while, just like there's no more 0auth. the open weights are just a marketing ploy.

[-]

NNN_Throwaway2@reddit

That would be my assumption. The fact that they were fan-faring availability on Arena was bizarre; normally they would just announce the release on their own platform.

[-]

charmander_cha@reddit

Preciso do modelo 9B urgentemente

[-]

dryadofelysium@reddit

first, they need to release the finished 3.7 models to cloud (only preview for now), after some time after that, they may open them

[-]

TheHiveFather@reddit

Its funny because Alibaba had some negative investor relations based on their spending, but still most maintain it as a buy position.. so all the moves aside, theres still something to be said about open models and not being afraid to back it up. In reality I think sharing open models if anything lets people who use this stuff in real ways day to day a better feel and eventually long term thats going to lead to more success within the development environment. Having a silo on your technology can only benefit short term, because eventually the need to expand, grow ans develop, you'll limit yourself.

That being said I thought after Gemma 4 releases we'd see a quicker response back. Could be a "dont release until we see something really worth showing off" or so one can hope. Either way, 2/5 models I run are Qwen and Im super impressed with them.

[-]

Porespellar@reddit (OP)

I agree with what you’re saying, and I can’t imagine that our “enthusiast” crowd running this stuff at home is going to hurt their bottom line too much, if anything we’re unpaid beta testers that often provide pretty detailed feedback that should be valuable to them.

[-]

TheHiveFather@reddit

I think you nailed it there, like just scraping reddit alone, people who run local are showing not only great end use feedback data set, but also integrations that they can continue to develop into their next model. I mean Gemini can only get so smart from your average user, and one could argue probably worse over time from general public use lol.. where as open weights in the hands of hobbyist/developers/researchers, people who want to find creative applications and push this technology.. its way more valuable to them.

(Alibaba if youre listening, trust us with your baby)

[-]

iMrParker@reddit

Don't do that. Don't give me hope

[-]

Ok_Librarian_7841@reddit

It ain't worse than gta 6