Reka Core, Flash, and Edge

Series of natively multimodal models processing text, images, video, and audio. Core (67B), Flash (21B), and Edge (7B) trained from scratch by a team of ~50 ex-DeepMind researchers. Reka Flash 3 (March 2025) is the latest open-weight variant.

Core approaches frontier models on human evaluations while Flash achieves SOTA for its compute class. AA Intelligence Index: 10 (Flash 3, 21B). Apache 2.0 (Flash 3). By the Reka Team.