Perplexity open-sources Decider 1.3: AI can find answers directly from the original text and pinpoint the exact location

Beating AI News Flash: Perplexity has released its next-generation decision model Decider v1.3 and the lightweight Decider Flash v1.3, built on Qwen3.8-27B and Qwen3.5-4B respectively. These models are purpose-built for tasks such as classification and scoring, and do not need to generate answers token by token like general-purpose large models.

The new version adds a Locate (source grounding) feature. For example, given a contract and asked to find the signatory, the model can directly return the corresponding name, its location in the source text, and a confidence score. Users can specify in natural language what to look for without pre-writing extraction rules. A single request can simultaneously perform classification, scoring, and information extraction.

Perplexity's internal testing shows that across 17 cross-domain information extraction tasks, Decider v1.3 scored 69.0, surpassing the comparison model Liquid d1's 43.3. Across 63 internal classification tasks, it achieved an average accuracy of 91.6%, 6.54 percentage points higher than the classification system the company currently uses in production.