Perplexity releases Decider v1.3 models with locate extraction

Perplexity released Decider v1.3 and Decider Flash v1.3 on Nov. 12 with a locate function that extracts matching passages from an input. The decision models return structured probabilities rather than free-form text, giving developers another option for classification, routing, moderation, ranking and evaluation.

The models are available through Perplexity’s Decisions API and OpenRouter. Perplexity says their downloadable weights are available on Hugging Face under the Apache 2.0 license.

The new locate question type returns verbatim text with start and end character offsets and a score. When it finds no match, it returns an empty span list. Locate can be combined with yes-or-no, multiple-choice and ordered-score questions about the same input, allowing an application to classify content and extract supporting text in one request. Perplexity’s API documentation says the feature is limited to the v1.3 models, accepts text state without images and has been evaluated on inputs up to 32,000 tokens, though the API accepts longer inputs.

Applications send content and one or more questions, then a Decider model returns probabilities that software can compare against thresholds. The Decider v1.3 model card and Flash model card display an Apache 2.0 license but say repository access requires authentication.

Perplexity says it trained the models by converting examples from hundreds of public datasets into a common format of state, questions, options and targets. The three-stage process uses generalist full fine-tuning, targeted low-rank adaptation and a vision-specific low-rank adaptation step. For Flash, the company used the larger model’s predicted probability distribution as the training target. It said this distillation method generalized better to held-out data than cross-entropy training on the same data.

V1.3 also uses bidirectional full-attention layers and an ‘isolated questions’ pattern, under which each question attends to the shared state and its own tokens. Perplexity said it chose that pattern for performance and answer stability. Locate adds a special token, span-prediction head and separate adapter while leaving the original backbone and decision head frozen. The company said it kept the locate adapter unmerged because merging it degraded ordinary decision requests.

Perplexity said it removed training samples that could overlap with its evaluations and conducted bias audits and removals. Those controls, the training corpus and the reported results have not been independently audited or replicated.

The company evaluated the models with external leaderboards and an internal suite of more than 200 zero-shot tasks. It reported 91.6% mean accuracy for Decider v1.3 across 63 classification tasks whose reference answers were relabeled by a frontier model, 6.54 percentage points above its production classifiers. Across 17 out-of-domain locate tasks, Perplexity reported average scores of 69.0 for Decider v1.3 and 63.6 for Flash, compared with 43.3 for Liquid d1, 41.5 for Clairic and 33.6 for GLiNER large v2.5; each task used its own primary metric.

Perplexity also tested answer stability on 372 multiple-choice questions. It said Decider v1.3 changed its semantic answer on 0.27% of questions under each of three option or label transformations, compared with 2.69% to 4.84% for Jev 1.14, and showed no changes when identical inputs were repeated.