Reka AIai lab
Also indexed as reka · reka-ai
Reka AI
This is a Model Atlas entity, not a ranked standalone lab. It is currently resolved under Moonvalley. A changing organization whose relationship with Moonvalley must be resolved before it is ranked independently.
Reka has deep multimodal research experience and is attempting to connect compact omni models with real-time video, devices, robotics, and physical-world data infrastructure.
Profile
Overview
Reka AI was founded in 2022 by researchers from DeepMind, Google Brain, Meta, and Baidu. It initially developed from-scratch multimodal language models including Reka Core, Flash, and Edge.
The company now focuses on physical AI, video understanding, world-model data, speech, and efficient edge deployment. Its merger with Moonvalley combines model research with video-generation infrastructure and large-scale visual data operations.
Notable contributions
- 01From-scratch compact multimodal model familyReka trained Core, Flash, and Edge as one family spanning frontier-scale service and small deployable multimodal inference.
- 02Video-first physical AI infrastructureThe current program joins world-model evaluation, egocentric data, video search, and edge models rather than treating video as an occasional input modality.