Hacker News Reader

they have reduced the token output by 20% and the benchmark scores have decreased by 10% of the original model.

The 20% output reduction is relative to R1, the 10% benchmark score reduction is relative to R1-0528.

It produces 60% fewer output tokens than R1-0528 and scores about 10% higher on their benchmark than R1.

So it's a way to turn R1-0528, which is better than R1 but slower, into a model that's worse than R1-0528 but better and faster than R1.