TheaterFire

Alibaba confirms they are committed to continuously open-sourcing new Qwen and Wan models

Posted by TKGaming_11@reddit | LocalLLaMA | View on Reddit | 79 comments

Alibaba confirms they are committed to continuously open-sourcing new Qwen and Wan models
Source: [https://x.com/ModelScope2022/status/2035652120729563290](https://x.com/ModelScope2022/status/2035652120729563290)

Reply to Post

79 Comments

vramkickedin@reddit

Looks like Wan 2.6 is now on openrouter "$0/M input tokens $0/M output tokens" [https://openrouter.ai/alibaba/wan-2.6](https://openrouter.ai/alibaba/wan-2.6) Interesting!
View on Reddit #81945009

wildcatshunter@reddit

https://preview.redd.it/frs7ze8n9jrg1.jpeg?width=374&format=pjpg&auto=webp&s=5c37f9eb6462e0e49861a9a3f003703648bfb021
View on Reddit #81718087

tom_mathews@reddit

Video gen locally is still rough — Wan 2.1 needs serious VRAM headroom to run without dropping frames or quantizing into mush. If they release the full series with proper small-model variants, that matters more than another Qwen iteration for most local setups.
View on Reddit #81691053

MerePotato@reddit

Half the talent already left though...
View on Reddit #81447079

Martialogrand@reddit

China competing and sabotaging AI bubble on US is something I’m enjoying too much
View on Reddit #81429543

Uncle___Marty@reddit

This makes me happy. Qwen3.5 has been so next level. Even the 0.8B was incredible.
View on Reddit #81317229

SufficientPie@reddit

> Even the 0.8B was incredible. for what?
View on Reddit #81337048

CATLLM@reddit

Ocr and translation. Requires good prompting but amazing a 0.8b model can do this
View on Reddit #81340769

SufficientPie@reddit

Oh, that makes sense. I forgot it has vision. Using it for OCR is better than OCR-specific models?
View on Reddit #81408893

CATLLM@reddit

Haven’t done a side by side comparison yet.
View on Reddit #81418653

specter800@reddit

I'm pretty new to all of this but how would a prompt improve those abilies? And what would that prompt look like?
View on Reddit #81342184

Yu2sama@reddit

Small models need more hand holding. A big model is smart enough (at times) to discern your intentions and do the work. A Small model will struggle with that, but with a good promt they can be very competitive. There is some paper about that, a Phi 2, Orca one and "Does model size matter?". If you are interested you can take a look at them.
View on Reddit #81366878

CATLLM@reddit

I had to be more explicit in what i want it to do. For example, i wanted to transcribe a screenshot that has a chinese text and then translate it into English. I would say, “here is a piece of text in traditional chinese. Transcribe the text in chinese then translate it into english.” Whereas with larger models i can just say “transcribe and translate this”.
View on Reddit #81342930

specter800@reddit

Oh I thought you were referring to a system prompt that would change the overall effectiveness.
View on Reddit #81347642

CATLLM@reddit

Yes that will work too
View on Reddit #81347803

specter800@reddit

Would you just make the system prompt similar what you suggested as the regular prompt and then never worry about it again?
View on Reddit #81347994

CATLLM@reddit

If that it’s only task then yes. But i don’t do that everyday so i just prompt it
View on Reddit #81352931

Admirable-Star7088@reddit

That's excellent news! I wonder though if their future models will suffer in terms of quality to some extent, given that several talented team members departed some time ago.
View on Reddit #81315310

Far-Low-4705@reddit

They can’t be any worse than what we currently have. But I would be more concerned about them falling behind other major open source models, especially with the loss of talent as you said
View on Reddit #81321901

IrisColt@reddit

Was Llama 4 noticeable worse than Llama 3.1?
View on Reddit #81336867

Far-Low-4705@reddit

if your comparing a 117b MOE to a 405b dense model... then yeah
View on Reddit #81349430

Mart-McUH@reddit

70B 3.3 and 3.1 (and probably 3.0 if you are ok with 8k context) are also better than Llama 4 Scout.
View on Reddit #81372091

Far-Low-4705@reddit

that is still a 70b DENSE model compared to a similarly sized sparse model... thats like comparing qwen 3.5 27b to qwen 3.5 35b and saying "look, bigger model worse!"
View on Reddit #81410216

lemondrops9@reddit

Llama scout was ok for chatting with, not much else use that I could find. Even then there are better models for that.
View on Reddit #81388572

Mountain_Patience231@reddit

yes, it is
View on Reddit #81346302

Ok_Warning2146@reddit

Will they open weight small models? If not, they are just another Chinese LLM company.
View on Reddit #81388838

Altruistic-Dust-2565@reddit

No, those Chinese characters just mean “More open-source models coming soon” and not specifying which series. Qwen I believe, but no guarantees on Wan as 2.5 and 2.6 are not open-source by far.
View on Reddit #81315396

lemondrops9@reddit

Ive seen other people posting that they will be opening 2.6
View on Reddit #81388647

coder543@reddit

But the slide title says: “Alibaba persists in open-sourcing the Qwen, Wan, and other series of models, advancing together with the ModelScope community.” The verb appears to imply a thing that has been going on and continues to go on. Not something that is over and done.
View on Reddit #81316339

Altruistic-Dust-2565@reddit

Yeah that's true. Didn't read the title. My bad. But still no official schedules yet for wan 2.5 and 2.6. Hopefully in the future.
View on Reddit #81350077

ANR2ME@reddit

May be they will open source the old Wan2.5 after they released API version of Wan 3 later 🤔
View on Reddit #81355560

goddess_peeler@reddit

It is not accurate to say that Wan open-sourcing is an ongoing thing.
View on Reddit #81316576

coder543@reddit

It may seem inaccurate, but that is clearly not their perspective, and the text at the bottom right says new models will be open sourced soon.
View on Reddit #81317370

goddess_peeler@reddit

I hope you’re right!
View on Reddit #81323445

toothpastespiders@reddit

Like when google releases a "new gemma" model. They're well aware that everyone wants and assumes it's a 9b/27b/etc size model when they start doing the "people who love open weight models should keep an eye on the gemma huggingface page! ~~~rocket ship emoji~~~ Then a couple more weeks of occasionally teasing, getting free PR talking about how great gemma and google are, and it's something like functiongemma or a tiny 270m gemma. It's especially important to consider when there's a language barrier and mistranslations might appear. I generally just assume that any tweet from a company promising something is at least 'some' kind of veiled lie. In the end twitter, for a company, is just a marketing platform and should be taken as seriously as a commercial but without any regulation on how much they can lie. Not saying it's the case here. But I think people should be a little more cynical about statements from companies and politicians on social media platforms. Lying through a carefully worded message that implies one thing while technically stating another is a time honored tradition.
View on Reddit #81326024

mikael110@reddit

As other have pointed out the slide titles is more specific, on top of that givne this is being tweeted by ModelScope, and is about an talk that occured in ModelScope DevCon you'd think they know what they are talking about, and that it's an endorsed message. If Qwen did not mean to imply that then the tweet would likely have been removed.
View on Reddit #81318890

lionellee77@reddit

The title of the slide specified Qwen and Wan. 
View on Reddit #81316537

ambient_temp_xeno@reddit

This is what I would expect. It wouldn't be smart to commit to "all future models" so people need to not mistranslate and get hopium.
View on Reddit #81316226

silenceimpaired@reddit

They could pull the typical thing… very small models for edge devices and very large models that require a data center. Hopefully not.
View on Reddit #81316353

Daniel_H212@reddit

Read the bottom left text
View on Reddit #81316300

hesperaux@reddit

Wan you say...? 🧐
View on Reddit #81377560

_derpiii_@reddit

I don't mean to sound cynical, just giving a reality check. Their commitment means nothing, because the CCP has ultimate authority. I'm not anti-China or anti-CCP. I'm just stating the power dynamics there. The CCP is literally above the law, which is hard for Westerners to grasp where law is above government.
View on Reddit #81365157

JayPSec@reddit

\`hard for Westerners to grasp where law is above government\` I think we're starting to get the idea...
View on Reddit #81374409

_derpiii_@reddit

> I think we're starting to get the idea... The misunderstanding comes with this nuance: us Westerners see it as corruption when the government is immune to the law. But from my understanding, it's completely different mindset in China where it's not considered corruption. It's just the way things are and the Chinese people wouldn't really want to change it. They have absolute trust in the authority of the CCP (I mean, well, as much as a human can have in that dynamic 😂). FWIW, I do believe America has more freedom and.. justice? But as the recent war suggests, we don't really have any control of the government and our situation isn't that much better. Our leaders are spending our tax dollars on Zionism, while China is building up its population and infrastructure.
View on Reddit #81375239

AnomalyNexus@reddit

Yay good news to start the week to
View on Reddit #81371769

Empty-Cake4502@reddit

Qwen's open source models are the best, thanks to qwen
View on Reddit #81369071

Illustrious-Lake2603@reddit

Can't wait for Qwen 3.5 Coder
View on Reddit #81316866

relmny@reddit

yeah, and even when I don't code! For some reason the "coder" qwen version can make a disappointing model (like qwen3-30b) become an extremely good model, for my general use case.
View on Reddit #81363435

jld1532@reddit

There is no way for-profit AI survives this, right? ChatGPT just announced ads in chat. Who is going to use that when LM Studio and powerful open weight models are free?
View on Reddit #81344338

mindwip@reddit

They release these small models and get great advertising, then host the better larger models. Both win and there apis generate money. First rule of tech companies is not to make money, but to get a huge user base no matter the cost, then figure our how to make money. There own app and api will make lots of money. Enough money to keep going lontpg term idk?
View on Reddit #81362908

Imaginary-Unit-3267@reddit

Normies who have no idea how AI even works but want to jump on the bandwagon anyway. People who don't own a GPU. Probably other groups.
View on Reddit #81355640

krigeta1@reddit

Cant wait for qwen image 2.0
View on Reddit #81359918

Impossible_Ground_15@reddit

This is great news
View on Reddit #81357912

c64z86@reddit

Awesome!! 🌠
View on Reddit #81356285

headfirst5376@reddit

need updated QWQ
View on Reddit #81355039

woct0rdho@reddit

Talk is cheap, show me the weights
View on Reddit #81353399

OneStrike255@reddit

This is good news! I think...
View on Reddit #81350854

vertigo235@reddit

So like, Will the next models be called "Qwen+" Or will the next models suck?
View on Reddit #81347773

the_real_druide67@reddit

Qwen has been quietly becoming my go-to for local inference on Apple Silicon. Ran Qwen3.5-35B on a Mac Mini M4 Pro 64GB : it pulls \~42 tok/s on standard prompts and still holds \~18 tok/s at 64k context. For comparison, most models of that size class choke hard past 16k. Alibaba open-sourcing aggressively is the best thing happening in local LLM right now. Meta started the race, DeepSeek proved you can do more with less, and Qwen is consistently shipping models that just *work* for real workloads.
View on Reddit #81347617

jreoka1@reddit

Thank God
View on Reddit #81346400

lionellee77@reddit

The left bottom mentioned: open source the full series of models, covers all sizes.
View on Reddit #81315208

sdmat@reddit

That's the key detail, thank you
View on Reddit #81345622

Spanky2k@reddit

If they wanted to sound convincing, they could have gone ahead and released open weights for Qwen Image 2 at the same time...
View on Reddit #81343495

foldl-li@reddit

Good news. ModelScope is co-founded by Alibaba, and this man is the driving force.
View on Reddit #81341487

DescriptionAsleep596@reddit

Well, I don't trust them
View on Reddit #81341135

ikkiho@reddit

alibaba open-sourcing makes total business sense tho. every dev building on qwen is a potential alibaba cloud customer, same way meta uses llama to drive their infra business. with deepseek and minimax both going open weights too, stopping now would mean losing developer mindshare overnight. the real competition isnt open vs closed anymore, its which open ecosystem captures the most users
View on Reddit #81340209

WithoutReason1729@reddit

Your post is getting popular and we just featured it on our Discord! [Come check it out!](https://discord.gg/PgFhZ8cnWW) 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.*
View on Reddit #81338735

Cuplike@reddit

Won't stop the obvious China haters here from talking about how in 2 more weeks every chinese company is gonna stop open-sourcing
View on Reddit #81336498

squirrelscrush@reddit

Common Alibaba W
View on Reddit #81329357

iamapizza@reddit

new models qwen
View on Reddit #81325599

Foreign_Risk_2031@reddit

They want to evaporate american investment money
View on Reddit #81322707

Noturavgrizzposter@reddit

When Z-Image-Edit?
View on Reddit #81321198

Significant_Fig_7581@reddit

They are the best ❤️
View on Reddit #81320378

DeepInEvil@reddit

Fu** openai and scam Altman
View on Reddit #81320042

the-final-frontiers@reddit

100% they will end up training them on chinese hardware which china will then dominate the gpu market.(few years) Which will be good for all of us to bring the prices down from the current madness.
View on Reddit #81319803

__JockY__@reddit

Qwen and MiniMax in one day. Hallelujah!
View on Reddit #81318382

No_Conversation9561@reddit

Wan hasn’t been open source for a while
View on Reddit #81317868

LegacyRemaster@reddit

waaaaaaaaaaaaaaaaaaaaaaaaaannnnnnnnnnnnnnnnnnnnnnnnnnnn
View on Reddit #81316148

dimaberlin@reddit

Qwen has been one of the strongest open families so far. If they expand both Qwen and Wan across multiple sizes, that’s a huge win for the community.
View on Reddit #81315854