Wow. While ClosedAI is burning through mega clusters of GPUs to release their non-reasoning 4.5o model, China is releasing amazing open source models.
QwQ 32B parameter model from Alibaba looks to be on par with R1 despite 1/20th the parameters.
pull down to refresh
I've been running it locally and so far its the only one I've tried that works half decent on a system with basic intel integrated graphics
Pretty amazing... any plans for how to utilize it or just playing with it for fun?
I'm surprised these Chinese companies are so openly sharing their secrets with us. Hard to argue we are ahead of them by any measure.
It seems every week a new breakthrough model is revealed it's so hard to keep up.
I'm experimenting with an orchestrator/MCP for Lightning.Pub, since chat bots are becoming the interface for everything else it seems intuitive that new users to Lightning/Bitcoin may be more comfortable with just using a chat based wallet. Obviously that has utility in automation as well.
I doubt China's motivations are entirely altruistic, OSS models provide a check on US proprietary models and there's precedent for using open source to play catch up.
That all said IMO things have stagnated since 4o, all the new stuff just seems to be better or recursive context servers in front of marginally improved models. The good news is I think anyone can discover use-cases with these commodity open source models because they're capable of most things you'd use a model for if inputs are conditioned right.
Ahhh very neat.
Meta has been surprisingly quiet... hopefully they have something up their sleeves or they might start to fall behind.
The frontier model space seems stressful competitive...
Yea its funny ollama exists because llama but i don't think anyone is using it to run llama anymore lol
https://xcancel.com/Alibaba_Qwen/status/1897361654763151544