Perplexity unveils decider v1.3, adding text span extraction capability

Perplexity has released its updated open-weight decision models, pplx-decider-v1.3 and pplx-decider-flash-v1.3. These models function as classifiers, returning probabilities for yes/no, multiple-choice, and score questions based on text or image inputs, rather than generating freeform text. The key enhancement in v1.3 is the addition of Locate, a feature that extracts precise text spans based on specific instructions or types, targeting tasks like NER and PII detection. The models are available under the Apache 2.0 license on Hugging Face and can be accessed via the Perplexity Decisions API and OpenRouter. The architecture adapts a generative LLM backbone, using a pruned LM head to output option probabilities. A notable architectural change in v1.3 is how questions interact; while the overall state sees all questions, each individual question now only interacts with the state and itself, preventing answer shifts seen in v1.2. For deployment, Perplexity recommends v1.3 for complex rubrics and Flash v1.3 for high-volume routing. The API endpoint is POST https://api.perplexity.ai/v1/decisions, with pricing set at $0.02 and $0.01 per 1M input tokens, respectively. Latency tests showed v1.3 handling a 243K-token request in 16 seconds, while Flash v1.3 completed it in 5 seconds.