DeepSeek V3 – The King is Back…For Free!
❤️ Check out Lambda here and sign up for their GPU Cloud:
Guide for using DeepSeek (R1) on Lambda (can be applied to DeepSeek V3 too, see links below):
📝 DeepSeek V3 (0324) is available here:
Try it online (note: they see your data, I prefer private, see below):
Paper:
How to run locally:
Ollama is probably the simplest way to run it – support is likely coming soon too! The previous version is available, keep an eye out for the V3 0324 version:
Ollama + 0324 versions might appear here:
Sources:
📝 My paper on simulations that look almost like reality is available for free here:
Or this is the orig. Nature Physics link with clickable citations:
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Benji Rabhan, B Shang, Christian Ahlin, Gordon Child, John Le, Juan Benet, Kyle Davis, Loyal Alchemist, Lukas Biewald, Michael Tedder, Owen Skarpness, Richard Sundvall, Steef, Taras Bobrovytsky, Thomas Krcmar, Tybie Fitzhugh, Ueli GallizziIf you wish to appear here or pick up other perks, click here:
My research:
X/Twitter:
Thumbnail design: Felícia Zsolnai-Fehér –
I’m glad these open models keep killing it
All ways fun geting notified from him.
The true Open AI
ironic isn’t it?
open AI should rename themselves to close AI
I love to hear that these models are becoming lighter to run
cant wait to have a model like that embedded on my smartphone or laptop fully offline
@@fiffy6572 this model is not lighter to run. It’s faster but larger, requiring more memory to run locally. It’s cheaper to run as a service but harder to run locally.
I agree that running on a phone would be great but deepseek v3 moves us further away from that goal, not closer.
These quants are craaaaaacked
I am glad they’re improving, to think this was started without much funding and
goes toe to toe w/ models that needs billions of $
I hope they release distilled models soon! I have R1 7b on my laptop, won’t be able to run a 271b param with anything less than a fully specced mac studio probably.
@@egriff38 671*b model
271b? you’ll likely need an H200, but if you’re referring to the 671b parameters in Deepseek, at least 4 H200 141GB cards
What about googles gemini 2.5 pro?
I like how none of reasoning models are capable of playing chess, while non-reasoning models sometimes play correctly.
It’s almost 700b thingies big 🙁 do they have any small ones brewing like with r1 or what
They figured out a way to extract the initial prompt from other AI’s.
Their new announcements on new features come weeks after other AI developers announce similar features.
And then Gemini 2.5 came along… But sadly, it’s not open source (not the time being at least) :/
WHAT A TIME TO BE ALIVE!!!
I’d like to know how it compares to claud 3.7 for programming tasks.
Wow another fantastic model…. waiting for the smaller model to run locally 😉
milkdrop visualizer ftw!
Waiting for a fireship video now
If this is the baseline for what’s free, the ceiling for proprietary AI just got raised.
What a fucking week it’s been so far. What a time to be alive it’s been so far!
Humans befitting from China as a rising power but not to dishonor US innovation. Both are befitting