They invented a hybrid attention design that drastically reduces the amount of memory needed for the KV cache at inference time. Like, dividing it by 10. And memory is a large part of the cost of inference.
This is a main part of the reason, yes. They actually innovated and did something that pushed the technology forward to be much more efficient, which we first saw with DeepSeek R1 for different reasons.
They invented a hybrid attention design that drastically reduces the amount of memory needed for the KV cache at inference time. Like, dividing it by 10. And memory is a large part of the cost of inference.
I see, I will look on some yt video who explain this in depth
This is a main part of the reason, yes. They actually innovated and did something that pushed the technology forward to be much more efficient, which we first saw with DeepSeek R1 for different reasons.