gokr.bsky.social
Loves programming, boats and badminton. Favorite language right now is Nim, at heart an old Smalltalker. Toying with AI and making a 2D action game slated for release 2025. Living north of Stockholm.
12 posts
13 followers
71 following
Conversation Starter
comment in response to
post
Just that a lot of folks mistake the distilled means "a smaller R1" when it is in fact "a Qwen fine tuned with synthetic data made by R1". It is however insane that a small Qwen can be tuned to that level. 2025 is going to be wild...
comment in response to
post
I note you use Ollama and deepseek-r1:7b - and I missed the end of the video so perhaps you explained that this is not Deepseek, but in fact a distilled (fine tuned) Qwen2.5. Cool nevertheless but don't want people to think anyone can run R1 or even V3 locally without FAT hardware :)
comment in response to
post
Sounds like Nim is what you want, perhaps you already looked at it? It has great automatic memory management and much more. I am using it for multiple things.
comment in response to
post
I also use it for building a game and various other private projects. It's a great under appreciated language.
comment in response to
post
beardygeeks.com - set that up a few months ago. The cup is really good quality, bought one from myself! :) Me and my friend even built custom hardware for the C64 so we could connect 8 joysticks...
comment in response to
post
Just for everyone else, this is Qwen, but fine tuned by R1. So it is not R1 itself. Nevertheless it is crazy interesting they can improve the smaller models this way!
comment in response to
post
Exactly which model? I presume you mean one of the distilled ones?
comment in response to
post
And I suspect smaller models for specific areas (like Qwen coder) will become more and more useful locally. And with open tooling like Cline and Aider, it's quite exciting. I use Cursor but Aider is awesome.
comment in response to
post
Yeah, I read someone bought an old workstation and smacked it with 768 Gb RAM etc, so yes, definitely doable but not for normal folks - yet... :)
comment in response to
post
But they also "distilled" (fine tuned) some well known smaller models by having Deepseek "teach" them. And those smaller models you can run on a modern laptop and they can match OpenAI midtier models like gpt4o, which is 🤩
comment in response to
post
Deepseek V3 and R1 are indeed fully open, you can download and use. Some companies are already offering V3 outside of Deepseek. And more importantly Deepseek is publishing papers on how they built them! But no, you can typically not run them on your own computer unless you have a smaller datacenter.