Perplexity Open-Sources pplx-decider v1.3 Decision Models: New Locate Span Extraction, 91.6% Mean Accuracy Across 63 Classification Tasks for the 26B Model, Apache 2.0, Free on OpenRouter Until November 19

Background and the problem

Agent systems make many small decisions: whether to route a request to a large or small model, whether a tool output contains a prompt injection, whether a search result is relevant. Many teams ask a general LLM to generate a text answer, which is slow, costly and inconsistent. Perplexity's Decider series takes another route and returns a probability for each option in a single forward pass. This release follows v1.2 in October.

Architecture and how it works

Decider prunes a generative model's output layer into a decision head that keeps only option tokens such as A, B and C, and reads probabilities at a special marker. It supports yes/no, multiple-choice and ordered-score questions, several per request. v1.3 makes the full-attention layers bidirectional and isolates questions from each other, which makes answers more stable. Training runs in three stages (generalist full fine-tuning, domain LoRA, vision LoRA), then KL distillation into the 4B model. The new Locate feature uses a separate adapter and span head to extract verbatim spans, loaded only for Locate requests so decision quality is unaffected.

Benchmarks and measured results

All figures are vendor-reported: 91.6% mean accuracy across 63 classification tasks, 6.54 points above Perplexity's production classifiers; nDCG@10 of 83.99 across 72 retrieval tasks, ahead of Qwen3-Reranker-8B and Jev 1.14; 88.16% mean accuracy across 8 languages (Flash 86.50%, Jev 1.14 85.41%); an extraction average of 69.0 versus 43.3 for Liquid d1. On one million real Decisions API requests, agreement with frontier LLM judges was 93.98% (v1.2: 93.68%). No third-party reproduction yet.

Getting started

Local use needs Python 3.12+ and about 49 GiB of GPU memory. Pull the weights with snapshot_download, load them with Decider.from_pretrained from pplx_decider, and call decide() with a state and questions; fast=True enables packed batching. The v1.3 request format is not compatible with v1.1 or v1.2, so old prompt builders cannot be reused. Use Flash for high-volume, latency-sensitive routing and the 26B model where mistakes are costly, and tune thresholds on your own labeled set.