Tasks and inputs
REKI-2B is listed for Image-text-to-text, Multimodal. It accepts image, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
REKI-2B is Kyutai's open 2.4B robot vision-language-action model with cheap register-token memory, pretrained only on public data.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights; no per-token price is published. Your cost is the GPU time you run it on.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Robot policies without memory oscillate. Reach for the cup, forget why, reach back. REKI-2B, from Kyutai with Inria and École des Ponts, adds memory almost for free. Thirty-two register tokens in the action expert summarise what the vision-language model saw at each frame and persist in the expert's cache across the last three frames, one second apart. Reading them adds under 1% to the time it takes to predict a chunk.
The payoff, by Kyutai's measures: oscillations between poses become about four times rarer, and success rises 1.5 points on LIBERO-Plus, 5.5 on RoboTwin clean scenes and 5.4 on RoboCasa365.
Everything in pretraining is public: real-robot teleoperation (Open X-Embodiment subset, DROID, BridgeData V2, AgiBot World and more), simulation, egocentric human-hand video and vision-language data. One billion frames in all, plus 12.2M vision-language samples, across 14 state layouts from single arms to humanoids. Training ran 200,000 steps on 32 H100s.
Kyutai reports the best scores among public-data models on RoboTwin Clean2Rand (65.8) and LIBERO-Plus (90.4).
The base appeared on 5 October 2026, with fine-tunes for LIBERO, RoboTwin (clean, and clean plus randomised) and RoboCasa365. Weights are CC BY-NC-SA 4.0, code is MIT. That rules out commercial deployment of the weights.
Real-robot use goes through LeRobot's async-inference gRPC protocol.
Start from the fine-tune closest to your hardware, say the LIBERO Franka model for a Panda arm. Reproduce its benchmark numbers in simulation before touching a real robot. Then fine-tune on fifty of your own demonstrations and count oscillation events per episode with memory on and off. That is the claim worth checking.
For a commercial-licence-friendlier alternative, look at FLUX 3 Action. Kyutai's speech models are on its provider page.
Compare the model on the work you need to complete.
REKI-2B is listed for Image-text-to-text, Multimodal. It accepts image, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under CC-BY-NC-SA-4.0 (weights); MIT (code). Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
NVIDIA
Alpamayo 2 Edge is NVIDIA's open 2B-class driving VLA that outputs trajectories with causal explanations, sized for DRIVE AGX Thor in-vehicle use.
shanghai-ai-lab
Sage-SciPerception is Shanghai AI Lab's set of projectors feeding crystals, spectra, biological sequences and biosignals into a language model's input.
inclusionAI
Ling 3.2 Flash is inclusionAI's 560B-A25B multimodal MoE with a 1M-token window, built for visual agents, 3D tasks and medical image reasoning.
Swiss AI
Apertus 1.5 8B is the Swiss AI Initiative's open multilingual model, now taking text, images and audio with a 262K-token context window.
black-forest-labs
FLUX 3 Action is Black Forest Labs' open 7B world action model for robots, predicting action chunks and future video from cameras, state and text.