Meet Kyutai REKI: The 2.4B VLA Model Bringing Efficient Memory to Robot Manipulation

Kyutai, in partnership with Inria and École des Ponts, has released REKI, a 2.4B parameter vision-language-action model designed for robot manipulation. The model is pretrained exclusively on public data and incorporates a low-cost memory mechanism that retains information from the last three frames, spaced one second apart. Across three out-of-distribution benchmarks, only models roughly twice its size outperformed it. The weights are currently available on Hugging Face.

The core specifications

• Size: 2.4B parameters, split between a 2.2B Qwen3.5-2B visual language model and a 0.2B flow-matching action expert. The memory module tracks the last three frames at one-second intervals.
• Hardware: Benchmarks were run on an NVIDIA H100. The system processes two camera inputs at 256×256 resolution to complete an action chunk in 21.8 milliseconds when compiled. Weights are distributed in BF16 format.
• Performance: It achieves the highest score among publicly pretrained VLAs on LIBERO-Plus and RoboTwin Clean2Rand.
• High score: 90.4% on LIBERO-Plus, the leading result in Kyutai’s comparison set.
• Low score: 38.4% on RoboCasa365, trailing Xiaomi-Robotics-1 which scored 57.4%.

What is REKI?

REKI functions as a vision-language-action model that translates camera feeds and natural language instructions into physical robot movements. It combines a Qwen3.5-2B backbone with a 200M-parameter action expert. This expert generates action chunks using flow matching, requiring 10 Euler steps for integration. Both robot state and flow time pass through adaptive RMSNorm. A small state MLP processes each specific robot state layout.

How the memory system works

Standard VLAs typically rely on the current frame alone. Previous methods attempted to include past frames by feeding them into the VLM, unrolling training sequences, or adding extra modules, each approach adding computational cost.

REKI embeds memory directly within the action expert. Before the action tokens for each frame are processed, the system inserts 32 learned register tokens. These tokens cross-attend to the encoded prompt and the current frame views. They also self-attend to the registers from earlier frames.

The registers do not read action tokens. Consequently, they depend solely on observations rather than flow time or robot state. They are calculated once per frame and stored in a rolling window cache.

The project documentation highlights the minimal overhead:

• Reading the cached registers adds 0.1 ms to the standard 21.8 ms action chunk.
• The cache consumes approximately 2 MiB of GPU memory per past frame.
• Training packs a three-frame window into a single batched pass.

The memory also resolves a specific failure mode. Without it, robotic arms frequently oscillate between two poses. On the RoboTwin benchmark, episodes showing oscillation dropped from 18.6% to 4.1% with memory enabled. Kyutai attributes this error to perceptual aliasing.

Benchmark results

REKI is pretrained once, then finetuned separately for each benchmark. The scores below are sourced from the model card and project page.

• LIBERO-Plus: 90.4%, surpassing Qwen-RobotManip (89.0%) and π0.5 (85.7%).
• RoboTwin Clean2Rand: 65.8%, second only to Qwen-RobotManip (68.5%).
• RoboCasa365: 38.4%, ahead of RLDX-1 (36.0%) and Qwen-RobotManip (35.9%).
• LIBERO (4 suites): 97.6%.
• RoboTwin 2.0 Easy / Hard: 85.9% and 84.6% respectively.

Pretraining provides the most significant boost for out-of-distribution tasks. It adds 13.8 points on Clean2Rand and 30.3 points under clutter conditions. The memory component contributes 1.5 points on LIBERO-Plus and 5.4 points on RoboCasa365.

Comparison with other VLAs

Notes: Clean2Rand means are calculated from the six conditions in Table 5 of the Qwen-RobotManip report. The Qwen report lists π0.5 at 84.4 on LIBERO-Plus.

Training data composition

The dataset mixture contains 1.0B frames and 12.2M vision-language samples. Real-robot teleoperation contributes 639.8M frames, simulation adds 193.9M, and egocentric human video provides 170.9M. Sources include DROID, BridgeData V2, AgiBot World, EgoDex and EgoVerse. Pretraining executed for 200,000 steps across 32 H100 GPUs. Kyutai retains each dataset’s native control convention rather than remapping actions.

Deployment details

Six checkpoints are listed: the pretrained base plus finetunes for LIBERO, RoboTwin (two variants), RoboCasa365 and RoboCasa-GR1. Real-robot deployment uses the LeRobot async-inference gRPC protocol. On a UR5 arm, REKI was finetuned for 3,000 steps, taking one hour, using 50 demonstrations per task.

The following code demonstrates the Python interface:

from reki import RekiPolicy

policy = RekiPolicy.from_pretrained("kyutai/reki-2b-libero", device="cuda")
policy.reset()  # clears the register cache at episode start
actions = policy.predict(images={"camera_top": img0, "camera_wrist_left": img1},
                         state=proprio, task="put the bowl on the plate")

What it means

The shift to a 2.4B parameter model suggests that high-performance robot manipulation no longer requires massive compute budgets. By integrating memory into the action expert rather than the visual backbone, the system reduces latency while improving stability against perceptual aliasing. The open availability of weights under a non-commercial license allows researchers to validate results without commercial constraints, though the model’s efficiency makes it a practical candidate for on-premise deployment.