On Oct 5 Reka Labs released a research preview of Rho-1, a 19B omni-reasoning model trained from scratch that unifies text, image, video and robot-action tokens in one network and one KV cache, with no tool calls or second model. Each transformer block has an understanding stream and a generation stream with shared attention, trained jointly with next-token prediction and flow matching. Reka reports base video generation at 0.79x real time (median) and a distilled 8-step variant (vs 99) returning a 5.3s clip in about a second. The checkpoint was trained on 320 H100s for three months. Research preview only; no public weights.
Key Takeaways
- ✓Scale: 19B parameters trained from scratch on just 320 H100s for three months
- ✓Speed: base video at 0.79x real time (median), ~6s to a watchable stream; distilled 99→8 denoising steps returns a 5.3s clip in ~1s
- ✓Architecture: dual expert streams (understanding/generation) per block with shared attention and one KV cache; next-token + flow-matching objectives
- ✓Limits: native video capped at 672x384; long-horizon drift, temporal grounding and edit stability are admitted weak points
- ✓Access: research preview only, no open weights or public API; partnerships via [email protected]

Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Reka Labs' Rho-1 (blog) is a 19B model trained from scratch that collapses the usual multimodal agent pipeline — a planner LLM handing off to image, video and robotics specialists — into one network with one KV cache. Text and commands are discrete tokens; image/video latents, robot actions and proprioception are continuous tokens. Each transformer block splits into an understanding stream and a generation stream with shared attention; the understanding stream emits a handoff token and the generation stream renders from the full accumulated state via a diffusion head. Training jointly optimizes next-token prediction and flow matching.
Reka publishes no standard benchmark scores, only internal speed numbers: base video at 0.79x real time (median) with a watchable stream in ~6s, and a distilled variant that cuts denoising from 99 to 8 steps and returns a 5.3s clip in about a second. Robotics is shown on LIBERO simulation demos without success-rate tables. Admitted limits: 672x384 native video, long-horizon structural drift, weak temporal grounding and brittle edits. The checkpoint used 320 H100s for three months. Rho-1 is a research preview with no open weights or public API; collaboration is via [email protected]. For developers the takeaway is architectural: a single model with shared state as an alternative to multi-model orchestration for robotics, simulation and interactive video.
This story is not about one specific product. Compare leading models side by side, or see which coding plan fits your team.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.