According to the language of the proposed bill, people who download AI models from China could face up to 20 years in jail, a million dollar fine, or both.
Base models are general purpose language models, mainly useful for AI researchers and people who want to build on top of them.
Instruct or chat models are chatbots. They are made by fine-tuning base models.
The V3 models linked by OP are Deepseek’s non-reasoning models, similar to Claude or ChatGPT4o. These are the “normal” chatbots that reply with whatever comes to their mind. Deepseek also has a reasoning model, R1. Such models take time to “think” before supplying their final answer; they tend to give better performance for stuff like math problems, at the cost of being slower to get the answer.
It should be mentioned that you probably won’t be able to run these models yourself unless you have a data center style rig with 4-5 GPUs. The Deepseek V3 and R1 models are chonky beasts. There are smaller “distilled” forms of R1 that are possible to run locally, though.
r1 is lightweight and optimized for local environments on a home PC. It’s supposed to be pretty good at programming and logic and kinda awkward at conversation.
v3 is powerful and meant to run on cloud servers. It’s supposed to make for some pretty convincing conversations.
That likely is one of the distilled versions I’m talking about. R1 is 720 GB, and wouldn’t even fit into memory on a normal computer. Heck, even the 1.58-bit quant is 131GB, which is outside the range of a normal desktop PC.
But I’m sure you know what version you’re running better than I do, so I’m not going to bother guessing.
You’re absolutely right, I wasn’t trying to get that in-depth, which is why I said “lightweight and optimized,” instead of “when using a distilled version” because that raises more questions than it answers. But I probably overgeneralized by making it a blanket statement like that.
Can you elaborate on the differences?
Base models are general purpose language models, mainly useful for AI researchers and people who want to build on top of them.
Instruct or chat models are chatbots. They are made by fine-tuning base models.
The V3 models linked by OP are Deepseek’s non-reasoning models, similar to Claude or ChatGPT4o. These are the “normal” chatbots that reply with whatever comes to their mind. Deepseek also has a reasoning model, R1. Such models take time to “think” before supplying their final answer; they tend to give better performance for stuff like math problems, at the cost of being slower to get the answer.
It should be mentioned that you probably won’t be able to run these models yourself unless you have a data center style rig with 4-5 GPUs. The Deepseek V3 and R1 models are chonky beasts. There are smaller “distilled” forms of R1 that are possible to run locally, though.
https://www.deepseekv3.com/en/download
I was assuming one was pre-trained and one wasn’t but don’t think that’s correct and don’t care enough to investigate further.
Is that website legit? I’ve only ever seen https://www.deepseek.com/
And I would personally recommend downloading from HuggingFace or Ollama
r1 is lightweight and optimized for local environments on a home PC. It’s supposed to be pretty good at programming and logic and kinda awkward at conversation.
v3 is powerful and meant to run on cloud servers. It’s supposed to make for some pretty convincing conversations.
R1 isn’t really runnable with a home rig. You might be able to run a distilled version of the model though!
Tell that to my home rig currently running the 671b model…
That likely is one of the distilled versions I’m talking about. R1 is 720 GB, and wouldn’t even fit into memory on a normal computer. Heck, even the 1.58-bit quant is 131GB, which is outside the range of a normal desktop PC.
But I’m sure you know what version you’re running better than I do, so I’m not going to bother guessing.
It’s not. I can run the 2.51bit quant
You must have a lot of memory, sounds like a lot of fun!
You’re absolutely right, I wasn’t trying to get that in-depth, which is why I said “lightweight and optimized,” instead of “when using a distilled version” because that raises more questions than it answers. But I probably overgeneralized by making it a blanket statement like that.