Introducing Reka Flash
- Source
- reka.ai
- Author
- Reka AI
- Date
Context
Reka AI's Reka Flash 3, a 21 billion parameter research preview released on March 10, 2025, writes out its reasoning before it answers, and developers can cap how long that reasoning runs. The release article, titled "Reasoning with Reka Flash 3", says the model wraps its thinking in reasoning tags. Forcing it to emit the closing tag after a chosen number of steps ends the thinking phase, and the model moves on to its answer.
Reka reports that this budget forcing still produces a reasonable output, and that Flash 3's score on AIME-2024 (a math competition , measured with consensus over 16 samples) keeps improving as the reasoning budget grows. Together these mean a developer can trade accuracy for shorter, cheaper responses and decide where to stop for their own task. The article shows the budget-versus-score results only for that one math benchmark, so how the tradeoff looks on other work is not established. The weights are downloadable under the Apache 2.0 license.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Checking sign-in…
Loading comments…