The opening demo, worth unpacking
On the evening of October 5, Reka released a preview of Rho-1 on its research blog. The demo starts with a generated lighthouse image; the model localizes an object in the scene, makes it move, changes the weather in the video, and then answers questions about the edit.
Reka's own description of this chain: each step reads from, and writes back to, the same shared state—there is no second model in the entire sequence, and no tool call at any point.
This is the core technical claim of the release. It compresses generation, simulation, reasoning, and action into one network, replacing what would otherwise be a four-stage pipeline stitched together by an agent.
Two expert streams: discrete and continuous tokens in one sequence
The architecture description is: the transformer block contains two expert streams. One handles language and visual understanding; the other generates images and video. The two share attention and context.
Tokens come in two kinds. Discrete tokens carry text, symbolic reasoning, and high-level instructions; continuous tokens carry image latents, plus video frames, robot actions, and proprioception. Putting symbolic reasoning and continuous numerical regression into the same autoregressive sequence is arguably the hardest part of this work. Language modeling uses cross-entropy on the next token, while predicting video frames requires regression on continuous values—two different objective functions trained jointly in one sequence. The paper gives no details on how gradients are balanced.
Tip: Proprioception refers to a robot's internal sense of its body position and joint states. Placing it in the same sequence as video frames means what the model sees and where it knows it is are the same kind of representation.
For robotics, Reka says the same model predicts future camera frames and outputs joint actions. The demo uses LIBERO simulation tasks. The release materials do not show Rho-1 controlling any physical robot.
Speed: 0.79x realtime; distilled variant produces 5.3 seconds in 1 second
All speed figures are company self-reported:
- The base model generates video at a median speed of 0.79x realtime, with streams starting around 6 seconds.
- A distilled variant compresses denoising steps from 99 to 8, producing 5.3-second clips in about 1 second in the company's tests.
- Structural drift during long video rollouts.
- Cross-video object anchoring remains unreliable.
- Targeted editing remains brittle across prompts.
- Native video output capped at 672x384.
- All architecture, speed, compute, and limitation figures come from Reka's official research blog and the October 5 release post; no technical report has been seen.
- 690K GPU-hours, 6.7x speedup, and the 16x compute gap are derived from stated card counts and durations, not company disclosures.
- 0.79x realtime and 5.3 seconds in 1 second are company self-tests with no independent evaluation.
- LIBERO is a simulation benchmark. The release materials show no physical robot control results.
By the latter figure, 5.3 seconds of video in 1 second implies roughly 5.3x realtime; relative to the base model's 0.79x, the distilled version pushes throughput about 6.7x faster. These two numbers describe two tiers of the same pipeline, not two separate models.
Native video output is capped at 672x384. This is below the video spec of any contemporary consumer device; Reka does not package it as a product capability and lists it directly in the limitations. It reads more like an internal research platform's self-use tier.
Training scale: trained from scratch on 320 H100 GPUs for about three months. Back-of-envelope, that is roughly 690,000 GPU-hours at 90 days of continuous full load—about sixteen times smaller than Beam's 11.2 million GPU-hours. This is not the same order of compute investment; keep that in mind when comparing capabilities.
The data path: mining action labels from internet video
The accepted hard problem in robot learning is data. Videos of humans doing things online number in billions of hours; robot data with sensor and action logs is far scarcer.
Reka's approach first trains an inverse dynamics model to extract control signals from ordinary internet video. Given a before-and-after pair of frames, the inverse dynamics model estimates what action caused the change; laying this over videos yields large-scale approximate action labels.
The weakness of this route is clear: the action labels are noisy. How a human arm moves in a video differs from a robot gripper in geometry, range of motion, and forces. Transferring from human video to a specific embodiment is the hard part—and where many demos that work in simulation fall apart on real hardware. Reka's release materials give no real-robot results and no list of tested embodiments.
Reka is not alone in this approach. Xiaomi's UMI line and Meta's open-source physical AI world models are also extracting control signals from human video in their own ways.
The company's own listed failure modes
Reka lists four, in blunt terms:
The company attributes these limitations mainly to training scale and data, and expects both to improve with scale—immediately followed by a self-imposed caveat: the preview does not establish that scaling will improve them.
No benchmark scores. No public pricing. No self-serve download. The release page invites teams working on embodied robotics, interactive simulation, and vision-action systems to contact them.
| Dimension | Rho-1 preview | |---|---| | Parameters | 19B, single network | | Modalities | Text, image, video, robot action | | Training compute | 320 H100s, ~3 months | | Robot evidence | LIBERO simulation, no real-robot demo | | Benchmark scores | Not reported | | License & weights | Not stated |
Who it competes with
Big players are already in this direction. Google DeepMind's Genie 3 is a realtime model generating navigable environments from text. Reka wants to put this kind of simulation, language reasoning, and robot action into one network.
The difference between the two routes is, for now, only at the architecture level. The existing demos do not show that Rho-1 can match systems purpose-built for interactive worlds or physical robot control. The gap between the architectural claim and reliable deployment is exactly what this preview makes visible.
On funding: Reka announced $110M in July 2025 led by NVIDIA and Snowflake; its 2023 round was led by DST Global Partners, with Radical Ventures as founding investor and Snowflake Ventures participating. This positions Rho-1 as research from a startup with a commercial product line, not an independent academic release.
CEO Dani Yogatama previously spent six years as a senior research scientist at Google DeepMind and is an associate professor of computer science at USC. His bet is an architectural one: a model that can predict what it will see next can also produce the action to take. At the level of physical deployment, that claim remains unverified by this material.