Reka Rho-1: One 19B Model Collapses Text, Video, and Robot Control Into a Single Network
Reka released a research preview of Rho-1 on October 5: a 19-billion-parameter model that unifies text, image understanding, video generation, and robot joint control into a single transformer with one KV cache. No external planners. No tool calls. No second model.
The technical report is titled “Rho-1: Collapsing the Multimodal Stack.” The name describes the thesis exactly. Most physical AI systems today are pipelines: a language model handles reasoning, a diffusion model handles generation, a controller translates plans into motor commands. Each handoff loses context and adds latency. Rho-1 tries to eliminate the handoffs.
The Architecture
Each transformer block in Rho-1 contains two streams. An understanding stream processes discrete tokens with next-token prediction. A generation stream handles continuous signals — images, video frames, robot joint trajectories — using flow matching. The streams share an attention layer, so the model’s understanding of a scene and its ability to generate from that scene are coupled at every layer.
Robot control is handled the same way. Rho-1 outputs joint angles as continuous tokens, not as calls to a separate controller. To address the scarcity of robot training data, Reka built an inverse dynamics model that extracts plausible control signals from ordinary internet video: watch a human arm reach for an object, infer the joint trajectory that would produce the same motion.
What the Numbers Say
Reka reports base video generation at 0.79x real time (median). A distilled variant that cuts denoising from 99 steps to 8 returns a 5.3-second clip in roughly one second on the evaluation setup. The model also works as a steerable world model: it generates continuous video and updates its trajectory mid-stream as new instructions arrive, without restarting.
The training budget is the headline number most researchers will focus on. 320 H100s. Three months. That is a fraction of the compute behind any current frontier model. Reka positions it as a proof of concept for the architectural approach, not a claim of frontier capability.
Current limitations are listed plainly: native video resolution tops out at 672x384. The model has short-horizon limits that make long robot tasks unreliable. No public weights. No API. Research preview only.
Why the Architecture Matters
The conventional argument against single-model physical AI is that specialised models optimise each component better than a joint model can. A dedicated vision backbone, a dedicated diffusion model, and a dedicated controller can each be trained on orders of magnitude more modality-specific data.
Rho-1’s response to that argument is that the handoffs between specialised models impose their own cost: each module sees only a filtered representation of what the previous module produced, losing the shared context that a unified model maintains throughout the interaction. Whether the shared context advantage compounds enough to overcome the specialisation disadvantage at scale is the experiment Rho-1 is running.
The 0.79x real-time video result suggests generation speed is not the bottleneck yet. The resolution ceiling and short-horizon limits suggest context saturation is. Scaling the architecture further — larger model, more modalities, longer sequences — will determine whether those limits are fundamental or engineering constraints.
No benchmark comparisons are published. Reka calls this a research preview, and the evaluation materials reflect that. The technical report is the primary artifact. An API or released weights have not been announced.
CEO Dani Yogatama built Reka from a team that came out of DeepMind and Google Brain. The lab was founded in 2022 and has been building multimodal systems since 2023, but has not previously published a physical AI result at this scope.