Reka AI, founded by over a dozen former Google DeepMind scientists, recently unveiled its first open-source model – Reka Flash3. This inference model, boasting 21 billion parameters, has garnered significant attention from both inside and outside the industry.
Despite its relatively modest 21 billion parameters, Reka Flash3 is, according to the developers, a general-purpose reasoning model trained from scratch. It has undergone supervised fine-tuning on synthetic and public datasets and been deeply optimized through Reinforcement Learning from both Model and Rules (RLOO).
This comprehensive approach has resulted in Reka Flash3 demonstrating impressive performance, even outperforming larger models like Command A and Gemma327B.
Even more exciting, Reka Flash3 is hailed as a leading open-source model of its size. This provides developers with a powerful and free tool to explore and build innovative applications.
Reka Flash3's capabilities extend far beyond simple text generation. It offers several compelling features:
- Strong Reasoning Capabilities: As a general-purpose reasoning model, Reka Flash3 excels at logical reasoning. Whether answering complex questions or performing multi-step inferences, it consistently demonstrates its intelligence.
- Low Latency and Device-Friendly Deployment: Reka Flash3 is designed for low latency and is highly suitable for on-device deployment. This means we may soon experience the convenience of this powerful model directly on our smartphones, tablets, and other devices. A boon for the efficiency-focused age!
- Core Driver of the Nexus Platform: Reka Flash3 is not just a standalone model; it's the core driver of Reka AI's new Nexus platform. Nexus aims to help organizations create and manage AI workers with native deep research capabilities, enabling them to browse the web, execute code, and analyze internal files including documents, images, videos, and audio.
- "Thinking" Mechanism and Budget Control: Reka Flash3 "thinks" before generating output, using specific tags to mark the beginning and end of its thought process. Interestingly, Reka AI also provides a “budget enforcement” mechanism, allowing users to force output after a certain number of steps, even if the thinking process isn't complete, often still yielding reasonable results.
While Reka Flash3 is remarkably capable, the developers acknowledge that as a relatively smaller model, it's not ideal for knowledge-intensive tasks. For such needs, Reka AI recommends combining Reka Flash3 with web search or other tools to broaden its knowledge base.
Model: https://top.aibase.com/tool/reka-flash-3
