Reka AI, a company co-founded by more than a dozen former Google DeepMind scientists, recently released its first open source model, Reka Flash3. Once this reasoning model with 21 billion parameters was launched, it attracted widespread attention from the industry.
Although the parameter scale of Reka Flash3 is only 21 billion, according to official reports, this model is a general reasoning model trained from scratch. It not only performs supervised fine-tuning on a variety of synthetic and public datasets, but also deeply optimizes through model and rules-based reinforcement learning (RLOO).

This "internal and external" training method has enabled Reka Flash3 to show amazing strength in performance, even surpassing models with larger parameters such as Command A and Gemma327B.
What's even more exciting is that Reka Flash3 is hailed as the leader among the open source models of the same scale. This means that developers finally have a powerful and free "hand weapon" that can explore and build various innovative applications more freely.
Reka Flash3's capabilities are much more than simple text generation. Behind it there are many exciting functional features:
First, as a general reasoning model, Reka Flash3 performs well in logical reasoning. Whether it is answering complex questions or conducting multi-step reasoning, it can show a "smart" side.
Secondly, the official emphasized that Reka Flash3 is a low-latency model and is very suitable for deployment on devices. This means that in the future, we may directly experience the convenient services brought by this powerful model on terminal devices such as mobile phones and tablets. This is simply a blessing in the era of "efficiency first"!
In addition, Reka Flash3 is not only an independent model, it is also the core driving force of Reka AI's new platform Nexus. The Nexus platform is designed to help organizations create and manage AI workers with native in-depth research capabilities that can browse web pages, execute code, and analyze internal files including documents, images, videos, and audio.
Reka Flash3 "thinks" before generating output and uses specific tags to mark the beginning and end of the thinking process. Interestingly, Reka AI also provides a "budget mandatory" mechanism that allows users to force their output after a certain number of steps of thinking, which usually produces reasonable results even if the thinking is not completed.
Despite its outstanding capabilities in Reka Flash3, the official also frankly pointed out that as a smaller model, it is not the best choice for knowledge-intensive tasks. For this type of demand, Reka AI recommends using Reka Flash3 with tools such as web search to give the model a broader source of knowledge.
Model: https://top.aibase.com/tool/reka-flash-3