Why OS isn't just about marketing for China
Posted by kaggleqrdl@reddit | LocalLLaMA | View on Reddit | 31 comments
A lot of people seem to think the OS releases was just a marketing gimic, a way to get into the US market due to fears of security.
But OS is always more then about that. It's about having leverage over standards and in this case, largely about GPU standards. By swamping the global market with powerful, cheap OS models they are rapidly becoming the standard.
When it comes time to new versions of hardware drivers, the question will be - does DeepSeek support it? Does Qwen support it?
These OS models give them a very powerful and compelling seat at the table.
colin_colout@reddit
In 2023 it really looked like meta was gonna win the OSS war.
...llama.cpp and ollama are literally named after their model. I always assumed they were looking to control the ecosystem, software stack, maybe even completion api spec...
But openai became the defacto completion api standard, and they really let the inference software ecosystem slip through their fingers.
...and llama4 was where they lost the edge on open** weight models.
** meta licensing is sus and quite restrictive
aeroumbria@reddit
BYW can we reclaim "llama" as a vendor agnostic term for language models, so I will never have to utter that awkward three letter acronym every again?
Environmental-Metal9@reddit
P.S.Y.?
_BreakingGood_@reddit
That's part of it, the complicated question of "Why is [big tech company] releasing this free and open source?" has a lot of answers, control of the ecosystem is one of them.
There are many other reasons too:
Recoil42@reddit
Another day on Reddit begging randos to stop framing everything that happens in China as expressly oppositional to the US.
Not everything is about you. Sometimes state and private actors do things for their own good, not because they're obsessed with stymying you.
Ylsid@reddit
What, and you also think the chip export restrictions aren't about disrupting Chinese capital too? Absolute lMao
Recoil42@reddit
Ylsid@reddit
Alright. Then I guess the CCP isn't at all doing what they can to subvert their opposition in a tech race, but the US is lol.
Recoil42@reddit
Learning about the last hundred years or so of history in US imperialism would do you well here.
Ylsid@reddit
Oh I'm well aware of their empire building atrocities. It's another reason I find it laughable to suggest the Chinese aren't taking every opportunity to do the same
Recoil42@reddit
Read my first comment again. Carefully.
Ylsid@reddit
It's still absolutely oppositional behavior. Of course the CCP is doing everything in their power to undermine US control of all AI technology, in much the same way the US is doing it back. Let's not act like they aren't.
Recoil42@reddit
_BreakingGood_@reddit
Wow, you really have to live in a bubble to not believe there is a massive AI war between the US and China right now. Did you know that China has export controls placed on key AI scientists?
Yes you read that right, China has placed export controls on certain Chinese citizens to prevent them from joining US AI companies.
Recoil42@reddit
You really almost had it there. Incredible.
Recoil42@reddit
Full-on brain-worms mode. Wonderful.
BobbyL2k@reddit
It took me too long to realize OS is “open source” and not operating system.
But yes, people build on top of open weights so they become defacto standards, but people also build on top state-of-the-art. So that’s why top companies are not totally replaced.
robertpro01@reddit
Yeah, OP should have said OSS or better FOSS
michaelsoft__binbows@reddit
OW...
threevi@reddit
Yeah, the proper acronym is OSS (Open Source Software) or FOSS (Free and Open Source Software) for that exact reason.
Regarding the reason why China makes their models open-source, I'd say it's simply because they aren't stupid. They're well aware of the reputation they have in the West, they have to know that if the only way to use their models was by connecting to their servers, nobody would use them out of privacy and security concerns. If they want to grow to rival US-based AI companies in terms of market share, they have to go out of their way to make their products look non-threatening.
livingbyvow2@reddit
Most likely, per a very good article written by Gwern, they are most likely simply commoditizing their complement .
What is pretty unique is that it used to be done at industry / companies level and now it's being done at the country level.
Some people are a bit too starry eyed when talking about US AI labs - it could very well be that it's like EVs - China initially looked late at the game, but a few years later they blew past everyone with BYD et allii. If the current AI wave is a bubble, their strategy may very well look prescient a few years from now.
Turbulent_Pin7635@reddit
It's not a bubble.
yopla@reddit
Same. I thought he was talking about Huawei's HarmonyOS or Xiaomi's HyperOS. Thought that was a weird sub to post but ok... Why not...
noctrex@reddit
Also lets not forget that China's Alibaba as of right now is the only country/company that has released Open Source models for the whole multimedia spectrum: text, image, video, voice, music. The Qwen and Wan families.
ac101m@reddit
Ultimately, LLMs (mathematically) are just matrices of numbers. For inference it really doesn't take much work to port an LLM to a new piece of hardware. If you have the weights and understand the structure of the model, it can be done relatively easily.
The current situation does highlight another issue with LLMs and AI investment though, and that's that if you expose your model for other people to use, you are also giving them access to the data they need to train their own model. If you ask Qwen3 235B 2507 who it is for instance, it will occasionally claim to be claude, probably because they trained it on synthetic data generated by claude.
So what these Chinese companies appear to be doing (at least partially) is training their own models against western ones, and then re-releasing them for cheap/free. Which begs the question, if this possible, then how can we continue to fund the development of new models? How does it make financial sense if they can just be stolen by your customers?
It's (unironically) the same problem that artists are facing at the moment, which would be funny if it weren't so bleak.
sautdepage@reddit
I really don't see the big problem, given that commercial LLMs are built on the fucking entirety of human knowledge, often copyrighted and/or with exclusive materials (ie. Google with their own user/video data, social media API deals, etc.)
To build something off all that and lock it down is way, way worse in my view than partially training on others models. Companies can provide services, but models should be open. No worries there's a ton of money selling AI services to business -- the cloud mostly runs on open-source software and hardware with commodity equivalents after all and have been doing just fine.
So as far as I'm concerned, proprietary models might as well not exist. For as long as I can, I will either running locally or running open-weight models via paid providers. If that doesn't work out, I will likely reject AI in my personal life.
At some point I need to put my money where my mouth is.
Durian881@reddit
I think you're too influenced by ongoing propaganda against China and US being the centre of the world. Google, Meta, Mistral, etc had been releasing open weight models before many of the Chinese companies came in.
fallingdowndizzyvr@reddit
Harmony OS?
prusswan@reddit
Technically they are not OS, just open weights, but anything to reduce reliance on cloud providers is a good thing.
jc2046@reddit
Interesting topic. I was wondering it myself yesterday... Like what´s the point in financial terms? This OS models costs a good bunch of millions to train and develop... Also, people here focusing in LLMs, but waht about image/video generation like WAN and Qwen image?.
AI_Tonic@reddit
that's not the direction support works : rather it would be "does it support deepseek and does it support qwen?"
so far, the answer is yes because there's obviously no support for anything - so far - in these chinese chipsets