Perplexity has launched Decider v1.3, a new iteration of its open-weight decision models. The system functions by adapting a generative LLM backbone into a classifier, where a linear decision head calculates probabilities for specific options rather than generating text. The v1.3 update introduces changes to the attention mechanism, ensuring that individual questions only interact with the shared state and themselves, while still processing all inputs in a single forward pass.
The new release includes a Locate feature designed for named entity recognition, PII detection, and extractive question answering. This feature utilizes a span head and a LoRA adapter, which remains unmerged to maintain decision accuracy. The model supports up to 128 questions per request. Developers can access the models via the API endpoint at api.perplexity.ai/v1/decisions using the identifiers pplx-decider-v1.3 and pplx-decider-flash-v1.3.
Pricing for the models is set at $0.02 and $0.01 per 1 million input tokens, while output tokens are provided free of charge. In internal latency tests, a request of approximately 243 000 tokens was processed in 16 seconds by the standard v1.3 model and 5 seconds by the Flash version. Perplexity recommends the standard v1.3 model for nuanced tasks and the Flash version for high-volume routing requirements.
