Log in Sign up
Opinion Taste-Bench From Microsoft and CityU Shows AI Agents Pick the Right Long-Term Move Only 60 Percent of the Time News Kyutai Releases REKI, a 2.4B-Parameter Robot Model With Built-In Memory That Punches Above Its Size AI Agents Google Open-Sources Harness2, a Framework That Lets an Agent Rewrite Its Own Scaffolding for Double-Digit Gains Apple Intelligence Apple Rewrites App Store Rules to Purge AI Nudify Apps, Putting Developer Accounts on the Line Opinion He Sold Everything and Rode a Bus 2,000 km to Marry an AI Farm Girl With 300,000 Followers Business Anthropic Pledges 5 Percent of Itself to a Public Interest Fund That Grows With Its Valuation AI Agents NVIDIA Open-Sources ASPIRE, the Robot Framework That Turns Failed Grasps Into a Growing Skill Library Coding Assistants PostgreSQL 19 Slips to November 19 After LLM-Assisted Code Review Turned Up Forty Defect Reports Opinion Taste-Bench From Microsoft and CityU Shows AI Agents Pick the Right Long-Term Move Only 60 Percent of the Time News Kyutai Releases REKI, a 2.4B-Parameter Robot Model With Built-In Memory That Punches Above Its Size AI Agents Google Open-Sources Harness2, a Framework That Lets an Agent Rewrite Its Own Scaffolding for Double-Digit Gains Apple Intelligence Apple Rewrites App Store Rules to Purge AI Nudify Apps, Putting Developer Accounts on the Line Opinion He Sold Everything and Rode a Bus 2,000 km to Marry an AI Farm Girl With 300,000 Followers Business Anthropic Pledges 5 Percent of Itself to a Public Interest Fund That Grows With Its Valuation AI Agents NVIDIA Open-Sources ASPIRE, the Robot Framework That Turns Failed Grasps Into a Growing Skill Library Coding Assistants PostgreSQL 19 Slips to November 19 After LLM-Assisted Code Review Turned Up Forty Defect Reports

Kyutai Releases REKI, a 2.4B-Parameter Robot Model With Built-In Memory That Punches Above Its Size

News · By: AIReport Admin · 1 hour ago · 👁 7
Kyutai Releases REKI, a 2.4B-Parameter Robot Model With Built-In Memory That Punches Above Its Size

French lab Kyutai, with Inria and Ecole des Ponts, open-sourced REKI, a 2.4B vision-language-action model pretrained exclusively on public data. The company-reported numbers show 90.4 percent on LIBERO-Plus and pose oscillations cut from 18.6 to 4.1 percent on RoboTwin thanks to a register-token memory that costs 0.1 ms per action chunk.

Kyutai, the French open-weights lab, has released REKI together with Inria and Ecole des Ponts, a 2.4-billion-parameter vision-language-action model for robot manipulation that is pretrained entirely on public data and shipped with everything attached: weights on Hugging Face under CC BY-NC-SA 4.0, code under MIT on GitHub, the training pipeline, and the deployment stack. The paper is on arXiv as 2610.22178, titled REKI: Registering Knowledge in VLAs.

The architecture is deliberately small. REKI pairs a Qwen3.5-2B vision-language backbone with a separate 200-million-parameter action expert that generates action chunks through flow matching over ten Euler steps, with robot state and flow time injected via adaptive RMSNorm and a small dedicated state MLP per robot body. On a single NVIDIA H100 the system computes an action chunk in a company-reported 21.8 milliseconds from two 256x256 cameras.

The memory design is the actual story. Most VLA models decide each moment from the current frame, so when a robot returns to an earlier step, putting a tool down, revisiting a part, it can oscillate between poses because nothing in the frame says where it has been. REKI fixes that inside the action expert rather than the backbone: 32 learned register tokens per frame cross-attend to the prompt and camera views, then self-attend to registers from earlier frames. The registers never read action tokens, so they depend only on observations, cache cheaply in a rolling window, and add a reported 0.1 milliseconds and roughly 2 MiB per past frame. The team says pose oscillations on RoboTwin drop from 18.6 to 4.1 percent when memory is enabled.

On benchmarks, all vendor-reported: REKI scores 90.4 percent on LIBERO-Plus, which stresses policies with perturbed cameras, lighting and layouts, and a 97.6 percent mean across the four LIBERO suites. Pretraining added a reported 13.8 points on RoboTwin Clean2Rand and 30.3 points under clutter. It is beaten on RoboCasa365, scoring 38.4 percent against 57.4 for Xiaomi-Robotics-1, a limit the team discloses itself.

Training ran a reported 200,000 steps on 32 H100 GPUs over 1.0 billion frames plus 12.2 million vision-language samples. On real hardware, a finetuned UR5 arm needed about an hour of training on 50 demonstrations per task to run the policy, though the released LIBERO card notes there is no real-robot or safety evaluation beyond those single-arm demonstrations.

The license deserves a second look. Weights are CC BY-NC-SA 4.0, free for research, off-limits for most commercial shipping, a choice that tempers the open-source celebration but is consistent with how Kyutai has handled earlier releases. The benchmarks are simulation, and the team says so plainly.

The bigger claim is the bet on architecture over scale. The VLA race has mostly been a brute-force contest of bigger backbones and more data, while REKI argues that cheap mechanisms, register tokens as compressed memory, well-curated public data, can buy reliability at a size a university lab or startup can actually run on one GPU. If that Pareto framing holds up under independent reproduction, the interesting consequence is not another leaderboard entry. It is that robot intelligence gets gated less on capital and more on ideas, which is exactly what an open-weights release is supposed to do.

Comments (0)

Log in to join the discussion

Log In

No comments yet