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.

It is always about the trade-off between those two parameters.

Of course an increase in both is the optimal, but a small sacrifice in performance/accuracy for being 200% faster is worth noting. Around 10% drop in accuracy for 200% speed-up, some would take it!

> Doesn't the deepseek licence completely forbid any use in the EU already?

Care to explain?

https://deepseeklicense.github.io/

https://github.com/deepseek-ai/DeepSeek-Coder/blob/main/LICE...

Sounds like a good use of "spare" time to me and not that different from many a lab I've been part of: someone gets a hunch, sets up an experiment to follow it, proves poor disproves whatever they were after, pulls down the experiment, rinse, repeat.

Yes and no.

Calling us a lab is not quite right, we are a consulting company.

But hacking is not just limited to in between placements, everybody has (at least) 2 days per month to do that, regardless of any work for customers.

Also, since AI is such a strategically important topic, we have a team that just works on AI stuff internally. That’s where R1T and R1T2 come from.

> The little 10% reasoning performance decrease seems worth it

We need about three orders of magnitude more tests to make these numbers meaningful.

Yes. If you look at the diagram that plots the performance vs the amount of output tokens, you can see that R1T2 uses about 1/3 of the output tokens that R1-0528 uses.

Keep in mind, the speed improvement doesn’t come from the model running any faster (it’s the exact same architecture as R1, after all) but from using less output tokens while still achieving very good results.