FT - China’s Alibaba shifts towards revenue over open-source AI

[-]

runner2012@reddit

you are right, it's more like theft, right?

[-]

CystralSkye@reddit

Ah yes reddit bites the hand that feeds it, the moment that their entitlements aren't met.

If you are localllama then stop begging for other people's compute? Train a model locally.

Like wtf is this level of entitlement.

[-]

Kubas_inko@reddit

Alibaba is not some self-funded dude/group in a basement. They are a multi-billion-dollar conglomerate.

There is literally no reason why we shouldn't be entitled, given their infinite budget compared to that of an average mortal.

[-]

CystralSkye@reddit

I can think of one reason, they are not affiliated with you unless you have some sort of business or government agreement with them.

And a lot more but I assume you are a thief? That just feels entitled to other people's property? In this case, property of a conglomerate associated with the chinese ccp?

Can't even call you a communist because, you know, you most likely aren't a citizen of china, so you wouldn't be able to communist your way to the property of a company in another country.

So I assume the proper term is literally thief.

[-]

Kubas_inko@reddit

What are you even talking about? To be a thief, you must take something away, remove access from the one being stolen from, which isn't happening in this case. You wouldn't download a car.

And China is not communist, other then it's political party having the word in the name.

[-]

There are plenty of people out there that does not care one bit about anyone or anything else than themselves and their family, do those people also have to pay for protection for these sense of entitlements that people have?

I mean it only makes sense if you look at it from a prospect of having conquest or theft. Can't get through business perspective? Must be acquired through conquest or theft.

I mean that is what people do already.

[-]

runner2012@reddit

no, it's not that.

What I meant is that, their product is actually built by stealing openai, anthropic, and other's western research.

it's like, when you buy a bootlegged movie, it's kinda silly that the price is almost on part with the original. But hey, if the market is willing to pay

[-]

CystralSkye@reddit

The same thing can be said about scraping the internet to make datasets which are used to make hobby models.

I don't understand this sub's mentality. Wants to local host stuff, but also want other people to provide them with models. It's peak irony.

[-]

Lesser-than@reddit

a subscribe to read article? Is there more information other than the qwen boss saying qwen will remain opensource "for now" in a tweet? or is it just more speculating over that tweet?

[-]

deepspace86@reddit

I guess nvidia is likely going to be the leading open source contributor going forward.

[-]

JMowery@reddit

They have no reason to do so in a significant/disruptive way. They want their partners to keep stuffing data centers full of insanely expensive GPUs. We will get open source models, but we're never going to get genuine attempts to disrupt the status quo.

[-]

kaggleqrdl@reddit

Except Anthropic goes TPU, Amazon has its own chips, and no doubt OpenAI will put the screws to nvidia or even comes up with their own hardware.

Nvidia needs a backup plan and OS is it.

[-]

JMowery@reddit

I don't understand what you are saying.

Are you implying that if Nvidia ends profitability in their enterprise division that open source and the every day consumers are suddenly going to make up the difference and save Nvidia from financial disaster by having all their consumers spend ~$20,000 - ~$50,000 on their products to handle local compute?

That's utter nonsense.

If current LLM enterprise dies for Nvidia... and say Google just takes over with TPU, then Nvidia just pivots to the next best thing, as they always do: robotics, self driving cars, lucrative government contracts (because you know tax dollars going to swoop in to save them).

It's just not even close to reality.

In fact, it's far, far more likely that AI itself suffers a total collapse altogether, and then you're just left with the likes of Apple, Google, and maybe a consolidation of a few of the other big players. And then AI will slowly rebuild to be actually useful and energy efficient instead of just throwing raw compute at the problem.

[-]

kaggleqrdl@reddit

consumers, lulz. You have no idea how many enterprises base their infra off of OSS models. Locking themselves into the frontier labs is waaaay to risky.

[-]

JMowery@reddit

You have no idea how many enterprises base their infra off of OSS models.

Show me the numbers, or stop talking.

[-]

kaggleqrdl@reddit

lulz. some people just don't want to be helped

[-]

Mochila-Mochila@reddit

In this scenario, nVidia would be stupid no to grab money where it can. And that includes local inference.

[-]

deepspace86@reddit

Nvidia has a vested interest in allowing people to run open models on their hardware. If the supply-side dries up, people will likely switch to AMD if gaming is the only (major) use case for consumer GPUs.

[-]

JMowery@reddit

Nvidia has a vested interest in allowing people to run open models on their hardware.

Yes they do. They are 100% regulatory. If the DOJ comes knocking and asks about open source, they now have a project to point to.

However, if you are implying that they care about the consumer? No, you are wrong. If you'd like to prove otherwise, give me some faction data / information on why that is.

Furthermore, your statement above has nothing to do with the original point.

Stay on point. The point was: "They have no reason to do so in a significant/disruptive way."

~20B open source models running on a local GPU does not matter.

If the supply-side dries up, people will likely switch to AMD if gaming is the only (major) use case for consumer GPUs.

No. People will NOT switch to AMD on a whim, because if that was the case, that would've happened back when AMD released the 7900 or 6900 (I forgot the exact model/year; I just know it was the last super competitive offering from AMD, and nothing has come close since then).

You also dismiss the vendor lock in factor. Everything runs on CUDA. Until there's an open standard that meets or beats CUDA, there's no concern from Nvidia.

A few thousand or 10s of thousands of nerds like us who push the limits of their GPUs to run local matters does not matter in the grand scheme of things.

30B local models are not disruptive. It's a fun side/hobby project for hardcore nerds. I like them. You like them. The overwhelming majority of people (90%+) are never going to own hardware this capable to run local models in a meaningful way to compete with your 1 trillion models.

We are not disruptive to the data center business, thus my original point is true, yours is off topic.

[-]

Guinness@reddit

This is really fucking stupid of them and I wonder if the Chinese government will have anything to say about it. Open models cause significant monetary damage to American companies.

Seed Dance 2.0 could take a massive chunk out of Hollywood too. If China wanted to fuck over the US, they should absolutely be releasing these tools as open source models.

Anthropic won’t make as much money if GLM 5.1 can do it for 1/3rd the cost.

[-]

ambient_temp_xeno@reddit

I think those of us without 1 tb of ram had better settle in and comfy for a while with Gemma 4 and Qwen 3.5/3.6

[-]

nuclearbananana@reddit

There's also - nanbeige - mistral - LFM - Arcee AI - LongCat - Nvidia

and many more labs. I wouldn't give up yet.

[-]

-Ellary-@reddit

Only Nvidia may drop something useful, but will they surpass Gemma 4?

Mistral with new limits (on training dataset) from EU, fails at all latest releases.
Their biggest MoE model is about same performance like Mistral Large 2 from 2024.
New small MoE not better than Mistral Small 3.2. And new Gemma 4 26b a4b butchers it.

[-]

cviperr33@reddit

Well they would probably start tightening things off, in a year imagine when we get our local claude mythos that is like ablitirated and uncensored , ppl are gonna target website or use it to create whatever chemicals they want lol

[-]

Alwaysragestillplay@reddit

Businesses rolling their own web services already use DAST and SAST scanners that have kept up with LLMs so far. Releasing a new generation of LLM isn't going to cause mass attacks anymore than the current gen already does. Decent scanners will integrate the new model as quickly as the hackers.

[-]

mxforest@reddit

I think Nvidia is our best chance going forward.

[-]

Geximus-therealone@reddit

maybe we all will have the opportunity to use open-source not completely for free, but for a paid subscription or even purchase

[-]

Mashic@reddit

No, we need them to be open so people can quantize and fine tune them for their needs.

[-]

ForsookComparison@reddit

Purchase model would be good.

I'd pay for Qwen models like I pay for video games on, say, GOG

[-]

Geximus-therealone@reddit

Maybe even the quality and responsibility will be a level higher...

[-]

QuackerEnte@reddit

yeah as if piracy isn't a thing lmao

[-]

QuackerEnte@reddit

I disagree with the top comment. I don't think distributed/decentralized training is the long term solution. I think what we truly need is AI model architectures capable of test time learning. Once you have (reliable enough) continual learning, you'll basically never need to heavily rely on big centralized tech ever again.

[-]

kaggleqrdl@reddit

and we will magic up these cool test time learning algs, right. bit of pixie dust, all we need

[-]

d70@reddit

Of course nobody who trains SOTA models makes money open sourcing stuff. They have to make money somehow.

[-]

brahh85@reddit

I dont buy these narratives, we never been better than in this moment.

[-]

vladlearns@reddit

it is true, and it makes sense, since the team lefy
I will repost here

"I think some of you don’t understand, but unless we figure out how to train on distributed infra, there’s no reason for big US companies to open source their models either. Mistral is great and all, but with what they have right now, they can’t compete

we need to coop and figure something out. it is not like hugging face and other stuff does not exist, but we can't pretend, that qwen, zai and minimax were not the biggest distillators and contributors"

[-]

_derpiii_@reddit

since the team left

Where did they leave to?

[-]

Emergency-Author-744@reddit

There is the distro psyche network that does distributed training: https://psyche.network/runs we will need more models trained like that to keep up

[-]

nomorebuttsplz@reddit

this is so cool. I wonder if there are hard limits to this approach. Like if a million 3090 owners all joined would there be a bottleneck such that it would be impossible to train a 1 trillion parameter model?

[-]

TwistedBrother@reddit

Heads are typically trained in parallel (good) per step (bad). I’m not aware of how one merges parameters successfully at this scale without deep model collapse. But I could be wrong.

[-]

SirReal14@reddit

A few years ago hivemind/petals was making big news, but the devs got hired by a different company and that development mostly seems to have stopped.
https://petals.dev/
https://github.com/learning-at-home/hivemind

[-]

vladlearns@reddit

let's goooo, I remember them from here https://www.reddit.com/r/LocalLLaMA/comments/1kmqmr8/nous_psyche_distributed_training_of_a_new_40b/

for the people, not the crown for freedom, not the king for our brothers, for our land, not for kings for liberty and the people no kings - only freedom For the cause, not the crown

[-]

Cool-Chemical-5629@reddit

it is true, and it makes sense, since the team left

Exactly 3 people left. Qwen organization on Huggingface currently has 192 members. Are you telling me that a team of nearly 200 people was held together by 3 individuals? If that's so, then maybe such team wasn't exactly going to last too long from the get go anyway.

[-]

vladlearns@reddit

4, not 3, but it does not matter:

Junyang Lin - the most important guy, their tech lead/architect of qwen series; Binyuan Hui – lead on qwen-code/r: he has basically built agent training pipe (pre- to post-training) + robotics; i tracked him down, he is at Meta, maybe there will be something from them soon - lets hope(there is 1 model already I reaaaally like, but it is not an llm); Bowen Yu - head of post-training - qwen-instruct, if you know; Kaixin Li – core on qwen 3.5/vl/coder too(vision-lang);

Lin had a strong advocacy for open weights, he was pushing the whole thing -> he left + other, who shared his ideology left, they started org restructuring, moved qwen out, isolating it = it was expected, that what I meant there

[-]

Cool-Chemical-5629@reddit

Last time I read about it there were 3 people leaving, but okay maybe there were more in the meantime. Still 4 people among hundreds is a drop of water in the sea and shouldn't shake the team to the point it falls apart. The real problem is the clear change of company's direction. In any case, I guess this also answers the question that's been circulating around here for a long time now regarding the catch of Chinese companies giving us stuff for free. It looks as though it was just a middle step to create awareness and now I guess they felt like it's time to milk the cow...

[-]

LegacyRemaster@reddit (OP)

yeah... bad news for us

[-]

Possible-Pirate9097@reddit

Maybe the real Qwen was the models we quantised along the way.

[-]

a_beautiful_rhind@reddit

They said they would keep open sourcing so I guess they lied. And now we know why they left the 397b off the list. We're never getting an actual big competitive model from them again.

[-]

LegacyRemaster@reddit (OP)

Chinese group Alibaba is reorienting its artificial intelligence strategy toward revenue-generating models, marking a potential shift from the open-source approach that contributed to the success of the Qwen family. The British newspaper "Financial Times" reports: Zhou Jingren, former chief technology officer of Alibaba Cloud, took over the leadership of the AI division following internal tensions over strategy and the departure of senior figures from the Qwen team.

The Chinese group now aims to strengthen its so-called "model-as-a-service" model, integrating AI tools into its e-commerce ecosystem and increasing the weight of proprietary models offered via the cloud. This move reflects a broader industry trend: the value is shifting from the pure performance of models to concrete applications, such as coding and autonomous agents, capable of generating ongoing use and revenue. According to the newspaper, internal concerns had grown over the high costs of open source and the lack of a clear monetization path, despite the strong support of Qwen from the global developer community. Alibaba will continue to publish advanced open source models in some areas, but the strategic priority now appears aligned with the group's cloud and revenue objectives.

[-]

Several-Tax31@reddit

Incredibly sad news, because qwen was the only one who develops small models for home pc's. The others are publishing big models, most of which are too big to run on pc's. And their models are really great.

[-]

HopePupal@reddit

qwen was the only one who develops small models for home pc's

this is… not even remotely true? we literally just got Nemotron 3 and Gemma 4. sure, Nemotron's not as good as Qwen, but Gemma's on par.

and on the very small side a few months ago we also got LFM 2.5 and Rnj-1 and FunctionGemma. and before that, Granite 4.

[-]