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.

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!

Care to explain?

https://deepseeklicense.github.io/

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

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.

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

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.