Reka Flash 3

Reka Flash 3 is a 21B-parameter, instruction-tuned language model developed by Reka, designed for general-purpose use. It performs strongly in chat, coding, instruction following, and function calling. With a 32K context window and reinforcement learning optimization (RLOO), it delivers competitive performance to proprietary models while maintaining a smaller footprint. Optimized for low-latency and on-device deployment, it supports efficient quantization (as small as 11GB at 4-bit precision) and uses explicit <reasoning> tags to indicate its internal thought process.