Reka AI Launches Rho-1 Omni-Model with 19 Billion Parameters

Reka AI's Rho-1 model features 19 billion parameters and utilizes 320 H100 GPUs for training, which took about three months to complete. The research preview of Rho-1 is scheduled for release on October 5, 2026, with the full Reka Core multimodal language model expected to ship in April 2024. Rho-1 uniquely processes all modalities as tokens within a single shared context window, eliminating the need for tool calls or external models. It can generate continuous video in real time and adapt to new instructions without restarting. Notably, the same model weights that predict camera images also control robot movements, demonstrating a seamless integration of various modalities. Reka AI has developed an inverse dynamics model that extracts control signals from ordinary internet videos, further enhancing the model's capabilities. This launch aligns with a broader trend in AI research focusing on world models, indicating a significant advancement in multimodal AI applications.