Reka Flash 3 is a 2.1 billion parameter general-purpose reasoning model trained from scratch, using synthetic and public datasets for supervised fine-tuning, combined with model-based and rule-based rewards for reinforcement learning. This model excels in low-latency and on-device deployment applications and possesses strong research capabilities. It is currently the best choice among similar open-source models and is suitable for various natural language processing tasks and application scenarios.