Reka Flash 3

Reka Flash 3 is a compact 21 billion parameter model released by Reka AI as an open-source research preview, positioned as a general-purpose assistant that handles chat, coding, instruction following, and function calling. Because of its relatively small footprint, Reka highlights its suitability for low-latency applications and on-device deployments, and describes it as the strongest model in its size category at the time of release. The model is also framed as a practical foundation for building domain-specific variants or a custom reasoning engine, giving developers flexibility to fine-tune or extend it for specialized workflows.

The training pipeline for Reka Flash 3 combines large-scale pretraining on a diverse mix of publicly accessible and synthetic data, followed by instruction tuning on curated high-quality datasets to sharpen task adherence. In the final stage, Reka applied reinforcement learning using REINFORCE Leave-One-Out (RLOO), pairing model-based and rule-based rewards to improve capabilities broadly rather than specializing for a single domain. The released research preview ships with a 32k context window, and Reka reports that the internal version continues to improve with additional training, with stated competitive performance against proprietary systems such as OpenAI o1-mini as a benchmark for the model class.