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.
"Due to the strict new guidelines of the EU AI Act that take effect on August 2nd 2025, we recommend that each R1T/R1T2 user in the EU either familiarizes themselves with these requirements and assess their compliance, or ceases using the model in the EU after August 1st, 2025."
Doesn't the deepseek licence completely forbid any use in the EU already? How can a german company legally build this in the first place (which they presumably did)?
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.
Fair point. More benchmarks are definitely good but I’m optimistic that they will show similar results.
Anecdotally, I can say that my personal experience with the model is in line with what the benchmarks claim: It’s a bit smarter than R1, a bit faster than R1, much faster than R1-0528, but not quite as smart. (Faster meaning less output tokens). For me, it’s at a sweet spot and I use it as daily driver.
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!
Also that “speed up” is actually hiding “less compute used” which is a proxy for cost. Assuming this is 200% faster purely because it needs less compute, that should mean it costs roughly 1/3 as much to run for a 10% decrease in quality of output.
Calling TNG a lab is a bit funny to me. It’s a consulting company that lets people hack on stuff between placements.
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.
OT: I love that German has a word for “yes and no”: jein.
So does English. Well, sorta.
Petition to make "nes" a word in english (yo doesn't really work...)
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.
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.
Yup, you can see it well on the graph here: https://venturebeat.com/wp-content/uploads/2025/07/Gu4d8kzWo...
From the hugginface model card:
"Due to the strict new guidelines of the EU AI Act that take effect on August 2nd 2025, we recommend that each R1T/R1T2 user in the EU either familiarizes themselves with these requirements and assess their compliance, or ceases using the model in the EU after August 1st, 2025."
Doesn't the deepseek licence completely forbid any use in the EU already? How can a german company legally build this in the first place (which they presumably did)?
> 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...
Probably a mix-up with the recently released Huawei model:
https://news.ycombinator.com/item?id=44441447
Is 200% a way to say *3 quicker ? The little 10% reasoning performance decrease seems worth it.
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.
> The little 10% reasoning performance decrease seems worth it
We need about three orders of magnitude more tests to make these numbers meaningful.
Fair point. More benchmarks are definitely good but I’m optimistic that they will show similar results.
Anecdotally, I can say that my personal experience with the model is in line with what the benchmarks claim: It’s a bit smarter than R1, a bit faster than R1, much faster than R1-0528, but not quite as smart. (Faster meaning less output tokens). For me, it’s at a sweet spot and I use it as daily driver.
If anybody wants to try it out, it’s up on chutes: https://chutes.ai/app/chute/4fa0c7f5-82f7-59d1-8996-661bb778...
tl;dr: faster but worse; i.e. on the pareto frontier.
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!
Also that “speed up” is actually hiding “less compute used” which is a proxy for cost. Assuming this is 200% faster purely because it needs less compute, that should mean it costs roughly 1/3 as much to run for a 10% decrease in quality of output.
↑