Most robot software is a relay race. A vision model labels the scene, a language model plans, a controller moves the joints, and each handoff loses information. Reka AI is betting on a single runner. On October 5, 2026, it previewed Rho-1, a 19-billion-parameter model that reasons in text, understands and generates video, and outputs robot actions, all as tokens in one network with no tool calls and no external models.
This is a research preview, and the coverage we reviewed includes no benchmarks, no license and no release of weights. So this post is about the design, which is interesting, and about what has to be shown before it counts as more than a promising direction.
TL;DR: the questions people are asking
| Question | Short answer |
|---|---|
| What is it? | A 19B omni-model: text, images, video and robot actions in one network. |
| Status? | Research preview. |
| Training? | About three months on 320 H100 GPUs. |
| Key idea? | Everything is tokens in one shared context; the same weights predict images and actions. |
| Data trick? | An inverse dynamics model extracts robot-style controls from internet video. |
| Benchmarks? | None reported in coverage we reviewed. |
| License / weights? | Not stated. |
| Usable today? | Not for production robotics. |
What "omni-model" means here
Reka uses the term omni-reasoning model, and describes a push toward "omni-world models" that can describe environments and also simulate and act within them. Three properties stand out.
One context for everything. Text, symbolic reasoning, image latents, video frames, robot actions and proprioception, the robot's sense of its own body position, all appear as tokens in a single sequence. There is no router calling a vision tool or a planning tool. The model reads and writes all of it.
Two token types. Reka says discrete tokens carry text, symbolic reasoning and high-level commands, while continuous tokens carry image latents, video frames, robot actions and proprioception. Mixing discrete and continuous outputs in one model is technically demanding, since language modeling and continuous prediction usually use different training objectives.
Real-time video and interruption. The model is described as generating continuous video in real time and responding to new instructions on the fly without restarting. For a robot, that implies it can revise its plan mid-motion rather than finishing a fixed sequence.
The idea that one network can imagine the future and act in it is a core theme of world-model research, which we explain in our world models guide. Rho-1 is a specific bet within that family: prediction and control share weights.
The same weights predict images and movements
Reka's most notable claim is that the same weights that predict camera images also drive robot movements. Why would that help? If the model learns to predict what the camera will see next, it must learn something about physics, objects and consequences. If the action head sits on top of the same representation, actions benefit from that understanding. This is the intuition behind recent "world action model" work, such as the scaling results we covered in Dyna-2.
The counterargument is that a model trained to predict pixels may waste capacity on irrelevant detail, and that debugging a single network is harder than debugging modules. Which effect dominates is an empirical question that needs results, not a press description.
The data problem and the inverse dynamics trick
Robot learning has a data shortage. There are billions of hours of video of people doing things and comparatively few hours of robots with matched sensor and action logs. Reka says it built an inverse dynamics model that extracts control signals from ordinary internet videos.
The idea: given two consecutive frames, an inverse dynamics model estimates what action caused the change. Run it over video, and you get approximate action labels at huge scale. Other teams use similar tricks, including Xiaomi's UMI-based work and Nvidia's open physical-AI world model.
The risks are obvious. Inferred actions are noisy. A human arm in a video is not a robot gripper, with different geometry, limits and forces. Transferring from human video to a specific robot is the hard part, and it is where many demos succeed in simulation and fail on hardware. Reka's coverage did not report real-robot results or the embodiments tested.
What is missing
Because this is a preview, the list of unknowns is long.
- Benchmarks. No numbers for reasoning, video quality or robot success rates.
- Robot evidence. Which robots, which tasks, how many trials, in simulation or real hardware.
- Latency. Real-time control needs fast inference. A 19B model generating video and actions has to run on a device or a nearby server with tight timing.
- Safety. A model that outputs physical actions needs safeguards, such as limits on force and speed and checks outside the model.
- License and access. No word on weights, API or terms.
- Compute claim context. 320 H100 GPUs for three months is modest compared with frontier training, which is notable, but it says nothing on its own about quality.
How to read a research preview
Previews are useful for tracking direction, and risky for planning. A few habits help.
- Separate the idea from the evidence. The architecture is plausible and interesting. The performance is unknown.
- Wait for a technical report with ablations, especially one comparing the unified model with a modular baseline on the same data.
- Look for real-world demos by independent labs, not only curated clips.
- Check the safety story before any physical deployment.
- Compare with open alternatives that you can actually run, such as Gemma on a small open robot or the models in our Gemini Robotics 2 coverage.
Why builders should still pay attention
If you do not work in robotics, there are still takeaways. The unified-token approach is spreading: models that read, write and act in one sequence are becoming normal in software agents too. The data trick, labeling raw video with an inferred action model, is a general pattern for turning abundant unlabeled data into supervised data. And Reka's modest compute budget, if the numbers hold, suggests that capable multimodal models do not always require frontier-scale spending, which is good news for smaller labs and for anyone hoping for more open models.
If you do work with robots, treat Rho-1 as one more candidate to evaluate when weights or an API appear. A short plan: define two or three tasks on your hardware, set success criteria, and test any new model on them in simulation first. Keep a human stop button regardless of the model.
Questions a technical report should answer
When Reka publishes more, a handful of questions will determine whether Rho-1 is a milestone or a demo. How does it compare with a modular baseline trained on the same data, since that isolates the benefit of unification? What is the action space and control frequency, and does the model run fast enough for closed-loop control? Which robots and tasks were tested, with how many trials, and how much of the evidence is from simulation? How does performance degrade when the camera, lighting or robot changes? What safeguards constrain physical output? And how much of the video generation quality is needed for control, versus being a side effect of training? Clear answers would turn an interesting preview into something other labs can build on.
Why the small training budget is worth noting
Three months on 320 H100 GPUs is roughly 230,000 GPU-days of compute. That is a serious but modest sum next to the largest frontier runs, which use many times more. If the model proves capable, it will support the argument that multimodal and embodied systems can be built by mid-sized labs, which widens who can contribute to robotics research. If it proves limited, the budget will explain why. Either way, the compute figure is a useful context for judging results when they appear. It also means a failed replication would be cheap to attempt, which is good for scientific scrutiny.
What this means for what you build or pay
Nothing to buy yet. The practical effect is on your reading list and roadmap: keep an eye out for a technical report, benchmark comparisons against modular stacks and an access announcement. If Reka publishes weights under a permissive license, a 19B omni-model that fits on modest hardware would be notable. Until then, this is a signal that unified world-action models are an active area, not a dependable tool.
Related reading
- What are world models? A complete guide
- Dyna-2: a world action model and robotics scaling law
- Nvidia Cosmos 3: open physical-AI world model
- Gemini Robotics 2: whole-body intelligence
- Xiaomi Robotics: 100K hours of UMI data
- Gemma 4 on an open duck mini robot
- Interfaze-1-Lite: an open model for deterministic tasks
Primary: Reka AI's Rho-1 research preview announcement (October 5, 2026) · The Decoder and MarkTechPost coverage
Details are accurate as of October 6, 2026 and rely on the vendor announcement and press coverage. No benchmarks, license or weights release were reported; treat capability claims as unverified until a technical report appears.