They’ve had distills before this, a more accurate title would be “Newest DeepSeek R1 distill runs on a single GPU like all the previous ones”.
Also it’s not accurate to say that a Qwen3 distill is the same as the DeepSeek R1 running in the datacenter - that one is still 85x larger than the Qwen3 distill.
What stands out about DeepSeek-R1-0528-Qwen3-8B is that it only requires a GPU with 40GB to 80GB of RAM to run
This is just inaccurate. It runs in 16GB of VRAM… because, you know, 8B parameters x 2 bytes (needed to store each parameter) = 16x10^9 bytes = 16GB…
It’s also just not the same thing at all. The distillations are not even remotely close to the 600+b parameters of the parent model. You can’t run Deepseek on your GPU at home. It’s the equivalent of buying your kid a powerwheels car. They’re not driving a car. They can’t drive a car at home.
Edit - I do run the distillations at home though and they’re fine for what I use them for. I’m not openai-pilled. I just hate when people say that you can run the same software as a massive datacenter privately.
It’s just sensationalism from a journalist who can’t even be bothered to multiply two numbers.
It’s gotta be a “distillate”, right? Not a “distill”. Verbing weirds language.
Most people say distilled model, distillate sounds right as well. The process is called distillation. I’ve just fried my brain on the local LLM subreddit because I was trying to get the transformers library working, probably why I phrased it like that.
I see
I’ve been running it on my laptop without a GPU
