Reka's 19B Rho-1 folds video generation and robot actions into one shared context
Reka released a research preview of Rho-1, a 19B-parameter model that generates text, images and video and emits robot actions in one network. Its robot results so far come from LIBERO simulation tasks, and physical control remains unproven.
The Engineer · Build desk
What happened
- Inside its transformer blocks, Rho-1 runs two expert streams, one for language and visual understanding and one for image and video generation, sharing attention and context.
- Reka says it trained the model from scratch on 320 H100 GPUs over roughly three months.
- Native video output in the preview is capped at 672 by 384 pixels.
- Reka announced a $110 million investment backed by NVIDIA and Snowflake in July 2025.
Why it matters
- What it limits constraint Reka rates object grounding across video as unreliable, and a manipulation policy has to track the object it acts on, so putting Rho-1 on hardware has to wait for that fix.
- What it costs cost A lab trying to reproduce the approach from scratch would need roughly 690,000 H100-hours of compute before it could compare results with Reka's.
- Who has to decide decision With no public pricing or self-serve download, a robotics team that wants to test Rho-1 on its own tasks has to start by contacting Reka directly.
Next story · Build · 6h ago
Gemini 4 Argon falls from 97.5% to 49.8% on ARC-AGI-3 once vendor features come offConfirmed2 publishers
The shared state is the strongest part of the engineering claim. In the familiar design, an agent links a chain of specialist models [6]. In Rho-1, Reka says, each turn reads from and writes to one shared state, with no second model or tool call in the sequence [5]. The demo conversation starts with a generated lighthouse image. It then locates an object in the scene, animates it, changes the weather in the video and answers a question about the edit [5]. Carrying one scene through all of those turns is clean design, and it is the part of the October 5th preview I would test first [1].
The robotics case reuses the same weights. The model predicts future camera frames and emits joint actions [7]. Reka's bet is that a system that predicts what it will see and produces the action can cut the handoffs between a world model, a planner and a robot policy [4]. Robot training data is scarce, so Reka built an inverse dynamics model that pulls control signals out of ordinary internet video, according to The Decoder [14]. Those action labels are inferred from footage. I'd want to see how far the labeller's errors carry into the joint actions the model emits. The announcement does not show Rho-1 controlling a physical robot [7]. Simulation is the right place to start. Nothing in LIBERO breaks when it falls over.
The speed figures come from Reka's own tests [12]. The base model generates video at a median 0.79 times real time, and a stream starts in roughly six seconds [8]. The Decoder describes the output as continuous real-time video that takes new instructions without restarting [19]. A median tells you about the middle run, and a live stream stalls on the slow ones. The distilled variant cuts the denoising path from 99 steps to eight and produced a 5.3-second clip in about one second [9]. That is about 12 times fewer steps [17], and output at roughly five seconds of video per second of compute [18]. For either figure to transfer, a buyer's prompts, clip lengths and serving hardware would have to match Reka's test setup.
Reka published its failure modes with the preview, and that is good practice. Long video rollouts drift in structure, and targeted edits remain brittle across prompts [10]. The company attributes its limitations largely to training scale and data, and expects both to improve as it scales them [20]. RuntimeWire noted that the preview has not established that scaling will improve them [21].
What to watch
- A Rho-1 demonstration on a physical robot, with action-rate and latency figures reported alongside the video ones.
- An independent measurement of the 0.79-times-real-time median and the distilled one-second clip.
- Whether a larger training run lifts the 672 by 384 output cap and fixes cross-video object grounding, as Reka expects from scaling.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption5
- Hype gap+30
- Incentives65
- Confidence55
Perspective Coverage
3 publishers- Builder
- Builder 57%
- Operator
- Operator 23%
- Investor
- Investor 20%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Reka released a research preview of Rho-1 on October 5th, a 19-billion-parameter model that the company says can understand and generate text, images and video, then produce robot actions within one network.
- [2]
Reka describes the architecture as two expert streams inside transformer blocks: one for language and visual understanding, another for generating images and video, with shared attention and context.
- [3]
Reka reports training the model from scratch on 320 H100 GPUs over roughly three months.
- [4]
Reka's thesis is that a system that predicts what it will see and produces the action to take can reduce the handoffs between a world model, a planner and a robot policy.
- [5]
The Rho-1 research post demonstrates a conversation that begins with a generated lighthouse image, locates an object in the scene, animates it, changes the weather in the video and answers a question about the edit. Reka says each turn reads from and writes to one shared state, with no second model or tool call in the sequence.
- [6]
Reka's announcement pitches a different architecture from the familiar chain of specialist models linked by an agent: Rho-1 keeps modalities in a shared context as a conversation unfolds.
- [7]
For robotics, Reka says the same model predicts future camera frames and emits joint actions. Its demonstration uses LIBERO simulation tasks; the announcement does not show Rho-1 controlling a physical robot.
- [8]
Reka says the base model generates video at a median 0.79 times real time, with a stream starting in roughly six seconds.
- [9]
A distilled variant, which reduces the denoising path from 99 steps to eight, produced a 5.3-second clip in about one second in the company's test.
- [10]
Reka lists constraints: long video rollouts can drift in structure, object grounding across video is still unreliable, and targeted edits remain brittle across prompts.
- [11]
Rho-1's native video output is capped at 672 by 384 pixels.
- [12]
The speed figures are company-reported; the launch materials do not provide an independent evaluation of the claims.
- [13]
Reka says Rho-1 is available as a research preview and invites teams working on embodied robotics, interactive simulation and vision-action systems to contact it. The launch post does not offer public pricing or a self-serve download.
- [14]
To work around scarce robot training data, Reka AI built an inverse dynamics model that pulls control signals from ordinary internet videos.
- [15]
In July 2025, Reka announced a $110 million investment backed by NVIDIA and Snowflake.
- [16]
Training Rho-1 took on the order of 691,200 H100-hours, assuming roughly three months of continuous use on 320 GPUs.
- [17]
The distilled variant uses about 12 times fewer denoising steps than the base path.
- [18]
In Reka's test the distilled variant produced roughly five seconds of video per second of compute.
- [19]
The model generates continuous video in real time and responds to new instructions on the fly without restarting.
- [20]
The company attributes the limitations largely to training scale and data, and expects both to improve as it scales them.
- [21]
The preview has not established that scaling will improve the listed limitations.
Sources
3 independent publishers whose own reporting we read for this story.
- dev.toReka's Rho-1: What Happens When One Model Replaces Your Multimodal Pipeline
1 article · October 6, 2026
- runtimewire.comReka releases a 19B model for video generation and robot control
1 article · October 5, 2026
- the-decoder.comReka AI's omni-model Rho-1 handles text, images, video, and robot control in a single model
1 article · October 5, 2026
Topics and entities
Topics
- World Models
- Video generation
- Natively multimodal models
- Embodied AI and Robotic Manipulation