Qwen 3.6 Max Preview just went live on the Qwen Chat website. It currently has the highest AA-Intelligence Index score among Chinese models (52) (Will it be open source?)

[-]

Few_Painter_5588@reddit

No, and they don't opensource the plus models either, so the 397B model.

[-]

MaxKruse96@reddit

397B is open weight. What are you talking about? The "plus" models arent new models, its just a software layer between user and inference engine to enhance them with extra serverside tools etc.

[-]

z_3454_pfk@reddit

3.5 plus is literally 3.5 397b. it’s literally in the model card on hf. 3.6 plus is 3.6 397b then which hasn’t been released.

[-]

Murinshin@reddit

Haven’t they only released one 3.6 model so far anyway?

[-]

Few_Painter_5588@reddit

No, 3.5 Plus (397B) was open weight. 3.6 Plus (397B) is not.

[-]

MaxKruse96@reddit

None of the 3.6 are except 35B. What kind of logic are we applying here

[-]

TheRealMasonMac@reddit

3.6 Plus was the first of their 3.6 series. The Plus model is 397B.

[-]

No_Doc_Here@reddit

let's hope the 122B actually comes :)

[-]

LatentSpacer@reddit

Notice how it’s compared against Opus 4.5 from Nov 2025 when we already have Opus 4.7.

[-]

Alex_1729@reddit

Interestingly, 4.5 was very similar to 4.6 performance, and a lot of people consider a 4.7 a non-nerfed 4.6.

[-]

LoveMind_AI@reddit

Honestly, Opus 4.5 was just about the best Anthropic ever got. The 1M context window was appreciated for the first couple weeks with 4.6 but things slid downhill fast.

[-]

Yeah they really peaked back in 4.5, which ironically made the claude models so popular they had to nerf what the users came for. They never gaf about non enterprise anyways, and I think there is a chance they are shooting their foot like what sam did after 4o

[-]

LoveMind_AI@reddit

At least Sam never really pretended to be an altruist. Dario’s out there acting like a moral crusader and it’s frankly gross. Talk about a company whose compass is spinning around and around.

[-]

muyuu@reddit

They're both insufferable tbh.

[-]

MuzafferMahi@reddit

Yep. And anthropic seems much more greedy, the anti- open source modelism and oss being “too dangerous” like fuck off

[-]

erizon@reddit

Benchmarks aside, the consensus in the community seems to be that Opus 4.7 is lazier/weaker than 4.6

[-]

No_Conversation9561@reddit

No, I believe 122B is the cutoff now.

[-]

Invent80@reddit

A 122b version of 3.6 based on what the 35b MoE is doing would be huge regardless. I can't believe how good the 35b model is. It's better than 3.5 122b for my use cases.

[-]

boredquince@reddit

that's sad :(

[-]

korino11@reddit

Veery strange. Becouse when 3/6 max was free. It always made a tons of mistakes. Reported that he have done the tasks at 100% When in reality project by Concept was done only on 30%.... 3/6 on 25% of his context started to forget the tasks... So yeah.. it is VERY strange. SOmethnig is abnormal here

[-]

Extra-Organization-6@reddit

if they release this open weight like they did with qwen 2.5, thats a game changer for self-hosted inference. a 52 on the intelligence index from a model you can actually run on your own infra is the real story here. most people dont care who has the highest benchmark score, they care what they can run without paying per token.

[-]

Thomas-Lore@reddit

They won't. They always keep Max closed. Hoepfully with Max out they will release the Plus now.

[-]

Extra-Organization-6@reddit

yeah Max has always been their closed flagship. if Plus drops with open weights at this benchmark level though, thats going to be wild for local inference. fingers crossed.

[-]

nullmove@reddit

It won't.

No local no care etc. but this feels like a good model. Worth reflecting that most people thought Qwen is done for when Junyang left, but turns out that Alibaba knows how to hire and they were getting someone because their Max models always sucked before. The new guy from GDM is probably an asshole if he actually compared old setup with intern's toy (pretty sure that was a fake quote?) but clearly he knows what he is doing to be producing results so quickly.

When they had closed off the 397B Qwen-3.6 model, I thought that was going to be their new Max. But now that they have pulled off an ever better Max so quickly, on one hand that makes me bullish on Qwen, but on the other it's now harder to swallow that they closed off the other one.

[-]

-dysangel-@reddit

3.6 came out pretty soon after 3.5. It seems more likely that 3.6 is based on training data that was being prepared while 3.5 was being baked, ie still brainchild of the old team. Do you have evidence to the contrary?

[-]

nullmove@reddit

Not really, no. But your assertion still strikes me as the more unlikely of the two, there was 7 months between Qwen3 (2507) and Qwen 3.5, that should be more than enough to give it their best shot for a major release. Training data preparation happens asynchronously in the background all the time anyway.

Obviously 3.6 is based on many good works done for 3.5, in particular the RL env scaling is probably a major factor which has to be built on a lot of painful infra work done by the old team. But I doubt ~1 month of extra data refinement was going to make a step change like we saw in 3.6.

And again, this is not an evidence but mere speculation, but given that the 3.5 27B dense model was virtually equivalent to the 122B MoE, and barely much worse than the big 397B one (with caveats, obviously bigger MoEs have more world knowledge), surely you have to say that 3.5 gen MoEs were quite underwhelming. I actually believe that MoE training is still hard with regards to router load balancing, and is something where frontier has superior expertise/secret sauce. I find it much more likely to believe that the new guy was able to improve MoE training regime with his imported expertise, than this just being about data works.

[-]

-dysangel-@reddit

Makes sense, I can see that.

The larger MoEs were underwhelming to me, but all the <40B models were state of the art for their size. 3.6 is clearly even better though.

[-]

tarruda@reddit

I haven't tested qwen 3.6 plus, but qwen 3.6 35b doesn't feel like a minor release. It is almost as if the 3.5 release was rushed and now they did the proper thing.

[-]

nullmove@reddit

Considering they took like 7 months before that release, if you set your feeling aside, do consider that it wasn't "rushed", just suboptimal MoE training that the new guy instantly fixed with his imported expertise within a month.

[-]

Long_comment_san@reddit

Qwen is doing an outstanding job of confusing with the names. If the model is more than 10% better on average, it's probably not 3.6 plus -> 3.6 max, it should probably be called 4.0. same with 3.5 which is a lot more powerful over 3.6 which confuses with a "small" number upgrade. Jeez, why bother with those number even

[-]

KnightNiwrem@reddit

I think plus/max is meant to indicate size of the model, rather than power of the model.

It's not the most confusing case yet. That prize goes to Gemini 3, where Flash is stronger than Pro in coding.

[-]

Long_comment_san@reddit

It's very confusing because for example 3.6 35b rivals 3.5 120b in many regards. That's the very definition of a case where you have to change the number significantly, it's a really large step in performance when a new model rivals a predecessor 3x it's size. Same with minimax actually, 2.5 doesn't hold a candle to 2.7 where the name transition implies maybe a 10% difference. GLM is no loser here either.

I mean it's just confusing. I don't have mental capacity to process what's the difference between qwen going from 3.5 to 3.6 and minimax going from 2.6 to 2.7 when both of them are "+0.1" in their naming which sounds like a simular level of improvement. Bruh

[-]

KnightNiwrem@reddit

The name transition does not imply a 10% difference. They are version numbers, not power levels, so this would be starting off from the entirely incorrect premise.

Typically, the first number in the version represents a major version, while the second number (the one after the period) represents a minor version. This is completely unrelated to performance, so you need to discard the idea of estimating improvements based on version number differences.

The version numbers typically are used to signal whether the product has undergone fundamental changes (i.e. major version change), or are merely refined versions without fundamental changes (i.e. minor version change).

[-]

mivog49274@reddit

Max is a well known series model in their ecosystem, since qwen 2.5 ?

it's the scaled up version of their generation training loop, I think it's about 1.5T in terms of size, never released, always beaten in benchmarks by their next generation of open models.

I don't know why "incremental" updates (.3,.4...) in all models releasing recently (GLM, GPT, Claude ect) delivers much powerful ones than in previous months, there seem to be a general acceleration since the end of 2025.

Really hope to have a Bonsai Qwen3.6-397B-A17B or Qwen3.6-122B-A10B, the 3.6 update was indeed quite a jump !

[-]

Steus_au@reddit

never say never

[-]

Lost_Promotion_3395@reddit

Solid score, but since it's a "Max Preview," they’ll likely keep the weights private to protect their lead before even thinking about an open-source release

[-]

No_Mango7658@reddit

Here's hopping plus was 122b and max is 397b!

[-]

Thomas-Lore@reddit

Plus was the 397b, they confirmed it pretty much on the day it was released.

[-]

No_Mango7658@reddit

Oh interesting! There's a chance we get that one

[-]

Middle_Bullfrog_6173@reddit

No, here's the blog, cleary states it's a proprietary model: https://qwen.ai/blog?id=qwen3.6-max-preview

[-]

LeTanLoc98@reddit

Why do they compare it to GLM 5.1?

GLM 5.1 is better than Qwen 3.6 Max.

They should compare it to GLM 5.

[-]

Cupakov@reddit

That’s precisely why they should be comparing it to GLM5.1

[-]

Dr_Me_123@reddit

Max never is.

[-]

DistanceSolar1449@reddit

They're no longer open sourcing the Plus model equivalents now, too.

They still haven't open sourced Qwen 3.6 397b even though Qwen 3.6 Plus released aged ago.

[-]

Cupakov@reddit

We’ve only seen one 3.6 release which was last week, and they said more’s coming

[-]

Cesar55142@reddit

Im still waiting for a 3.6 Omni that i can self host

[-]

VoiceApprehensive893@reddit

gguf when

[-]

Pakobbix@reddit

Max Models were never available to us, so I doubt it.

But I'm curious on how many parameters this model has. Plus is the 397B, so if the 397B 3.6.. 600-700B?

[-]

Kamal965@reddit

Probably larger. Qwen3 Max was over 1T parameters, according to Qwen/Alibaba, although AFAIK they never mentioned the active parameter count: "The Qwen3-Max model has over 1 trillion parameters and was pretrained on 36 trillion tokens."

Realistically, Qwen3 Max probably had ~90 to 100B parameters active if they maintained a similar sparsity as Qwen3 30B-A3B and Qwen3 236B-A22B.

Qwen3.6 Max... honestly, who the fuck knows lol. Qwen3.5 397B-A17B has a 4.28% sparsity, whereas the 122B-A10B and 35B-A3B models have an ~8.20% and ~8.57% sparsity respectively. If I (baselessly!) assume the parameter count of Qwen3.6 Max was scaled up from Qwen3 Max in a similar ratio as the other Qwen3.5 models were compared to Qwen3... well there's actually no reasonable guess to be made as there's no consistency in how they scaled up the parameter count of the models. Qwen3.5 397B-A17B is ~69% larger than the 235B model, and Qwen3.5 35B-A3B is ~16.67% larger than the 35B-A3B models (and I'm flat out ignoring the 122B-A10B model lol). If we use both as an upper and lower bound on a baseline assumption that Qwen3 Max was exactly 1T parameters, that would mean Qwen3.6 Max is anywhere between 16.67% to 69% larger, so somewhere between 1.167T to 1.69T parameters.

As you can see, that's a lot of assumptions made. It's assumptions all the way down. William of Ockham would be ashamed of me.

[-]