3 ms·
No one knows the size of o1. The only hint was a paper that suggested it was 200B parameters. Meanwhile, DeepSeek R1 is known to be 671B parameters because it
by minimaxir 2y ago
No one knows the size of o1. The only hint was a paper that suggested it was 200B parameters.
Meanwhile, DeepSeek R1 is known to be 671B parameters because it is open-source.
- sigmoid10 2y agoR1 is a mixture model with only 37B active params. So while it's definitely expensive to train, it's rather light on compute during inference. All you really need is lots of memory.