AI model · by Reka

Reka Flash 3

Reka Flash 3 is Reka's model for building apps that need quick responses or deployment directly on devices.

  • 66K tokens context
  • Open weights · Apache 2.0
  • Checked against Reka's page 19 Oct 2026

What Reka Flash 3 is for

Reka designed Flash 3 as a base for apps that depend on fast response times or running locally on hardware. Visitors can test it on Reka Space, and it ships in a format compatible with Llama, so any library that supports Llama can load it. Reka says Flash 3 also powers Nexus, its service where businesses build and run AI workers.

What Reka says it does well

  • Can stop its thinking early via budget forcing and still produce a reasonable output.
  • Released in a format that is easy to deploy with Llama-compatible tools.

Where to use it

  • Try it out on Reka Space.
  • Released in a Llama-compatible format, so any Llama-compatible library can run it, such as Hugging Face transformers or vLLM.
  • Powers Nexus, Reka's platform for organizations to create and manage AI workers.

Limits Reka mentions

  • As a smaller model, it is not the best at knowledge-intensive tasks; the maker recommends pairing it with web search for those.
  • Often thinks in English when prompted in non-English languages, which can affect output quality in those languages.
  • Has not undergone extensive alignment or persona training.