Rho-1
modelYour notes
Reka's 19B "omni-reasoning" model trained from scratch, released on October 5, 2026 as a research preview. One network reads and writes text, images, video and robot actions. Reka presents it as a replacement for agentic pipelines in which a planner model hands work to separate image, video and detection models: in Rho-1 every turn reads from and writes to one KV cache, which Reka describes as a persistent world state. In the launch demo the model draws an image, puts a bounding box on it, animates it into a video, edits the video and then describes the edit, all in one conversation with no tool calls.
The architecture is symmetric. Text and commands are discrete tokens; image latents, video frames, robot actions and proprioception are continuous tokens. Each transformer block splits its weights into two expert streams that share attention over the same KV cache: an understanding stream with a next-token text head, and a generation stream that denoises latents with a flow-matching diffusion head. The model is trained end to end with next-token prediction on the discrete sequences and flow matching on the continuous ones. Reka reports training on 320 H100 GPUs for three months, and attributes most of the failure modes it lists to that modest compute and data budget.
Reka gives no standard benchmark scores. Its own measurements: the base model generates video at a median 0.79× real time, with a watchable stream after about 6 seconds, and returned a first clip in 7.0 seconds in a race against an illustrative 13.8-second multi-agent pipeline. A distilled Rho-1 Flash cuts denoising from 99 steps to 8 and returned a 5.3-second clip in about one second, with prompt edits rendered in 1.11–2.01 seconds. The post also shows the model as a steerable real-time world model and as a world-language-action policy on LIBERO simulation tasks, where it predicts future frames and the seven action channels from the same latent state. It pairs with Reka's Inverse Dynamics Model, which infers actions from unlabelled video so web video can serve as action-labelled pretraining data. Reka lists long-horizon structural drift, no grounding across video yet, brittle editing consistency and a native video cap of 672×384 as limitations. No weights or paper have been released; access is by request through Reka.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| Rho-1 | 19B | Research preview; base model, 0.79x real-time video generation (median) |
| Rho-1 Flash | — | Distilled variant, 8 denoising steps instead of 99; described, not separately released |