Perplexity open-sources AI brains that decide things instead of chatting

Perplexity has released version 1.3 of its open-source Decider models, pplx-decider-v1.3 and the smaller pplx-decider-flash-v1.3. These models are designed to make structured decisions instead of generating text, answering yes/no, multiple-choice, or scoring questions about text or images. The main upgrade is a new "Locate" feature that can pull out specific quotes or spans of text based on a user's instructions. Both models are available for free under the Apache 2.0 license on Hugging Face and can be accessed via Perplexity's API platform.

The models are built on different-sized versions of Qwen backbones, with the main model requiring a hefty 49GB of GPU memory. Perplexity claims it outperforms a competitor's model on most of its internal benchmarks, scoring over 91% average accuracy on a broad set of classification tasks. However, its new 'Locate' feature is its weakest point, and the smaller 'Flash' model lags behind in specialized areas like finance and law. This release provides developers with a powerful, freely-available tool for adding structured reasoning to applications without relying on text generation.