REKI
A 2.4B-parameter vision-language-action (VLA) model released by Kyutai for robot manipulation. It is pretrained solely on public data, incorporates a built-in low-cost recent-memory mechanism, and is released with open weights.
Timeline (last 90 days)
French AI lab KainTai has released the REKI vision-language-action model, with only 2.4 billion parameters, and all weights have been open-sourced.
REKI was pre-trained entirely on public data.
REKI embeds a cheap memory of its recent past within the model, rather than attaching a large external video understanding module.
On three out-of-distribution benchmarks, only models with roughly twice REKI's parameter count were able to surpass REKI.
Kyutai, together with Inria and École des Ponts, released REKI, a 2.4-billion-parameter vision-language-action (VLA) model specifically aimed at robotic manipulation tasks.
All of REKI's model weights are already publicly available.
REKI's positioning is not to pursue first place on SOTA leaderboards, but to prove that under limited parameters and public data conditions, VLA can also possess usable manipulation capability and generalization.
REKI's actual performance still needs more third-party reproduction to verify, and there is still a gap between benchmark numbers and real deployment.
The original article does not list the specific names of these three out-of-distribution benchmarks or the list of comparison models one by one.
All Facts (9)
REKI's positioning is not to pursue first place on SOTA leaderboards, but to prove that under limited parameters and public data conditions, VLA can also possess usable manipulation capability and generalization.
50%UnverifiedREKI was pre-trained entirely on public data.
50%UnverifiedREKI embeds a cheap memory of its recent past within the model, rather than attaching a large external video understanding module.
50%UnverifiedOn three out-of-distribution benchmarks, only models with roughly twice REKI's parameter count were able to surpass REKI.
50%UnverifiedThe original article does not list the specific names of these three out-of-distribution benchmarks or the list of comparison models one by one.
50%UnverifiedREKI's actual performance still needs more third-party reproduction to verify, and there is still a gap between benchmark numbers and real deployment.
50%UnverifiedFrench AI lab KainTai has released the REKI vision-language-action model, with only 2.4 billion parameters, and all weights have been open-sourced.
50%UnverifiedAll of REKI's model weights are already publicly available.
50%UnverifiedKyutai, together with Inria and École des Ponts, released REKI, a 2.4-billion-parameter vision-language-action (VLA) model specifically aimed at robotic manipulation tasks.
50%