Reka releases Rho-1, a 19B omni-model trained from scratch on 320 H100s in 3 months
RekaAILabs · x · 2026-10-05
Reka AI released a research preview of Rho-1, a 19B omni-reasoning model trained from scratch on roughly 320 H100s over about 3 months — a tiny fraction of frontier-scale compute.
- Core idea: most AI today is an agentic pipeline where a central model delegates to modality specialists; Rho-1 instead collapses the stack, unifying text, vision, and robotic actions as tokens in a single context window.
- One network understands and generates text, images, and video, reasons over them, and takes actions — demoed by rendering a lighthouse image after deliberate composition planning, boxing an object on it, then animating it into a drone-approach video.
- The team frames it as its first shot at proving unified omni world models are the future; the blog includes full transcripts and unedited session rollouts.
Related event: Reka unveils Rho-1, a 19B unified any-to-any model trained on 320 H100s(8 posts)→
More from Multimodal
- Dev combines Meta's DINOv3 and SAM 3.1 to track individual instruments in video — MaziyarPanahi · 2026-10-06
- NVIDIA open-sources Lyra 2.0: turn any photo into an explorable 3D world — Promptmethus · 2026-10-06
- Grok Build turns a single Roadster photo into an interactive 3D demo with a fly mode — XFreeze · 2026-10-06
- How to escape the glossy 'AI oil painting' look: ComfyUI user seeks naive 19th-century Russian style — CommunicationNo8431 · 2026-10-06
- Open-source video models are essential for indie filmmakers, argues one creator — taherdhanera · 2026-10-06
- HyperFrames Studio launches: a video editor built for agents, co-editing on one timeline — toolstelegraph · 2026-10-06