More of Silicon Valley is building on free Chinese AI
Posted by buppermint@reddit | LocalLLaMA | View on Reddit | 29 comments
Posted by buppermint@reddit | LocalLLaMA | View on Reddit | 29 comments
Velvet-Thunder-RIP@reddit
I mean its nice to get something from them if we can. If they want to give advanced models away for free there will be some applications that benefit us more than them.
red780@reddit
I think the Chinese view it as an economic war... Free AI hurts the West much more than China (manufacturing). China also loves to control human thought - why not try for AI.
Hopefully we've all over-built on AI hardware so much that when they try to take over Taiwan it's less of a disaster.
croninsiglos@reddit
They aren't wrong. I'm far more likely to build on a performant model I can both control, easily fine tune, and self-host. This doesn't mean I don't still use frontier models, but if I'm building something stable then I need a stable base. We need more American releases like GPT-OSS or Gemma which can compete in this space at a reliable cadence without providers fearing it'll cannibalize their cloud offerings if it's "too good".
Turbulent_Pin7635@reddit
-p-e-w-@reddit
Add to that that the line between so-called frontier models and Chinese models is really blurring now, with several weak releases from OpenAI and Anthropic while Chinese models continue to leapfrog forward. Kimi K2 Thinking is actually better than GPT 5.1 and Opus 4.5 on some tasks, and is far ahead when it comes to making quality conversation.
tavirabon@reddit
Gemini 3 Pro was a massive leap forward. GPT5 was a massive leap forward in efficiency and 5.1 pushed capability forward a lot. Opus 4.5 was a massive leap forward in efficiency (especially considering their own (perf/$):($/token) history) and also a step forward in capability. No frontier AI has been having weak releases. Further, Kimi K2 has 1/4th the context size as 3 Pro and every model will have a "better at x task" for the foreseeable future. The lines are not blurred, you're just squinting really hard.
TheRealMasonMac@reddit
Gemini 3 is worse than 2.5 Pro.
VampiroMedicado@reddit
I like R1 for planning but it’s a bit neurotic
twack3r@reddit
You really should give MiniMax M2 a go on that list.
__JockY__@reddit
Sure, but let me know when I can run Gemini 3 Pro on my K2 rig.
Capable_Site_2891@reddit
It's interesting because they aren't really Chinese models.
They're absolutely brilliant distills of American models, produced by Chinese companies.
There is censorship in the API layer, but not really in the models.
TomorrowsLogic57@reddit
Agreed! Small models specialized for single or very limited use cases will be a force to reckon with in the years to come.
Idc if they are built on open source Chinese models or US models, this shit is happening regardless.
SlowFail2433@reddit
If its apache 2.0 or MIT license then you don’t have to interact with the company anyway
LostMitosis@reddit
The fear is gone.
Dry_Yam_4597@reddit
China is killing it. In AI and robotics.
Piyh@reddit
Especially killing in Xinjiang
bene_42069@reddit
While the social discrimination does exist for Uyghurs in China, the "genocide" and "ethnical cleansing" is a media over-exaggeration. It's kinda more like Papuans in Indonesia. How about you pay attention to actual genocides like in Palestine or Sudan instead?
polytique@reddit
While unrelated to the subject of LLMs, it’s more than social discrimination. Uyghurs are being mass deported to concentration camps, the women sterilized, and all of them forced to work in factories near the camps.
No_Conversation9561@reddit
Korean company. Most of the models we use are chinese models.
strangeanswers@reddit
we use open source chinese embedding models for our RAG pipelines. american big tech company
synn89@reddit
Oof. That's going to be hard to beat.
procgen@reddit
Isn’t all of that good for Silicon Valley? Why should they complain when these models are free?
LocoMod@reddit
It’s not true. Unfortunately.
exaknight21@reddit
Bro, it’s America, we thrive on exploiting cheap stuff - China giving out free models, to me, is the biggest F U to capitalism. Say what you will.
eposnix@reddit
America exploiting cheap labor from China is a tale as old as time.
equitymans@reddit
Uh ohhhhh :)
LocoMod@reddit
Nobody is using the startups mentioned in the article. Ask anyone working for a profitable established business. There is OpenAI, Anthropic and Google. That’s the only game in enterprise town. Any jank startup can fine tune to save pennies, but the reality is that they are passing the cost to the consumer because it takes more tokens to solve problems with open weight models (fine tuned or not) via multi turn prompting than it is to just pay the closed source tax and get your solution in one or two turns. The people valuing these startups have to park money somewhere. They have obscene amounts of capital they can gamble. 8 billion? That’s nothing. Someone made a gamble and backed it up with capital. It’s only worth that much to the investors that already paid the price for it. So yea, that’s the value to the investors holding the bag. And it is their job to produce articles like this to hype the investment so they can sell the bag to some schmuck that falls for this crap. And a lot of them will. Poor schmucks. Filthy rich schmucks. No matter. Someone paid for it and they will do everything they can to convince the potential buyers that bag is worth more. But it’s only worth is the price that was paid. Nothing more. If I was rich and put in a bid to buy Reflection for 3 trillion dollars and I put up the capital, that’s what it’s worth to me. Now my job is to convince you to buy the bag for 5 trillion. And know this: I will poison the media with these articles to convince you that’s the real worth.
Reflection is worth nothing to the rest of the world. The bag holder is desperate. Don’t fall for it.
Arcosim@reddit
Hard not to when they're releasing banger after banger completely open source. I'm getting better results with Z-Image than with any close source genAI model.
kinkvoid@reddit
not surprised!