this post was submitted on 02 Feb 2025
52 points (79.5% liked)
Technology
61346 readers
2862 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
In fairness, unless you have about 800GB of VRAM/HBM you're not running true Deepseek yet. The smaller models are Llama or Qwen distilled from Deepseek R1.
I'm really hoping Deepseek releases smaller models that I can fit on a 16GB GPU and try at home.
Well, honestly: I have this kind of computational power at my university, and we are in dire need of a locally hosted LLM for a project, so at least for me as a researcher, its really really cool to have that.
Lucky you! I need to check my university's current GPU power but sadly my thesis won't be needing that kind of horsepower, so I won't be able to give it a try unless I pay AWS or someone else for it on my own dime.
Sure but i can run the decensored quants of those distils on my pc, I dont need to even open the article to know that openai isnt going to allow me to do that and so isnt really relevant.
Qwen 2.5 is already amazing for a 14B, so I don’t see how deepseek can improve that much with a new base model, even if they continue train it.
Perhaps we need to meet in the middle, and have quad channel APUs like Strix Halo become more common, and maybe release like 40-80GB MoE models. Perhaps bitnet ones?
Or design them for asynchronous inference.
I just don’t see how 20B-ish models can perform like one orders of magnitude bigger without a paradigm shift.
I use 14b and it's certainly great for my modest highschool physics and python (to help the kids) needs, but for party games and such it's a drag its pop culture stops at mid 2023
Thing is, there are a lot of free APIs for 30B-70B class models.
Self hosting is great of course, and if 14B does the job then it does the job.
Intriguingly, there's reason to believe the R1 distills are nowhere close to their peak performance. In the R1 paper they say that the models are released as proofs of concept of the power of distillation, and the performance can probably be improved by doing an additional reinforcement learning step (like what was done to turn V3 into R1). But they said they basically couldn't be bothered to do it and are leaving it for the community to try.
2025 is going to be very interesting in this space.
nVidia's new Digits workstation, while expensive from a consumer standpoint, should be a great tool for local inferencing research. $3000 for 128GB isn't a crazy amount for a university or other researcher to spend, especially when you look at the price of the 5090.
Why would you buy a single use behemoth when you can buy a strix halo 128GB that can work as an actual tablet/laptop and have all the functionality of the behemoth?! while supporting decades of legacy x86 software. Truly wondering why anyone would buy that NVIDIA thing other than pure ignorance and marketing says NV is the AI company.
Dense models that would fit in 100-ish GB like mistral large would be really slow on that box, and there isn't a SOTA MoE for that size yet.
So, unless you need tons of batching/parallel requests, its... kinda neither here nor there?
As someone else said, the calculus changes with cheaper Strix Halo boxes (assuming those mini PCs are under $3K).