Sorry if you guys get so overwhelmed with deepseek submissions these days. This will be my one and only in the next time. It is cool to have an anti-weight to all these pay models.
Personally I don't get sick of it. There's a lot of hype around Deepseek specifically rn, but to run SOTA or near SOTA models locally is a huge deal, even if it's slow.
The issue is that this article is conflating (as do many, many articles about the topic) the distilled versions of R1 (basically llama/qwen reasoning finetunes) with the real thing. We are not even talking about quantized versions of R1 here, so it's not quite accurate to say you're running R1 here.
Hey the model from https://ollama.com/library/deepseek-r1 is used.
This is not the real thing? Just 671b is the real thing or what are you going to tell me?
19 comments
[ 2.6 ms ] story [ 54.1 ms ] threadhttps://ollama.com/library/deepseek-r1
This includes the full model and additionally several distills.
Saw this in the article
>I would not recommend running this on your main system. Unless you like unnecessary risks.
Using hosted versions where host collects data or using a unknown software that runs the model is the risk.
> Why Ollama? Because it makes running large language models actually easy.
> If it doesn’t work, fix your system. That’s not my problem.
what an off-putting start