Perplexity Releases Decider V1.3 Models with Text-Span Extraction

Perplexity released Decider v1.3 and Decider Flash v1.3 on November 12, 2026, adding verbatim text-span extraction and downloadable Apache-2.0 weights. The company reports 91.6% mean accuracy for the larger model across 63 query-classification tasks with references independently relabeled by a frontier LLM, 6.54 percentage points above its production classifiers.

The new Locate feature returns exact matching text with start and end character positions for named-entity recognition, personal-information detection and extractive question answering; no match produces an empty list. Both models are also available through OpenRouter and Perplexity’s Decisions API.

Decider returns probabilities for yes/no questions, custom choices or ordered scores instead of generating prose. An application can ask several questions about the same text, JSON or images and use the resulting numbers for classification, request routing, moderation, filtering, output evaluation and reranking. These are tasks that often need a decision threshold rather than free-form generation.

Architecture and Training Changes

Decider adapts generative backbones by replacing next-token output with a decision head read at a marker token. The head starts from only the base-model rows associated with answer-option tokens. In v1.3, Perplexity makes full-attention layers bidirectional and isolates questions from direct attention to one another, while allowing the shared input state to attend to the questions. That replaces v1.2’s arrangement, in which an answer could change depending on what else was asked, and is designed to make answers more stable.

Training converts hundreds of public datasets into a shared state, question, options and target format. Generalist fine-tuning excludes benchmark training splits, and Perplexity says it also removes samples that could overlap its evaluations. Later domain and vision specialization uses LoRA, a method that freezes pretrained weights while training injected rank-decomposition matrices. Perplexity uses weighted adapter merges to control drift and averages weights from three runs for the text stages.

Flash is subsequently distilled from the larger model, using KL divergence to reduce the difference between its output probabilities and the teacher’s rather than learning from gold labels alone. The Decider v1.3 model card specifies a Qwen3.8-27B backbone and roughly 49 GiB of BF16 weights; the Decider Flash v1.3 card specifies Qwen3.5-4B and roughly 8.5 GiB. Self-hosting either model requires a CUDA GPU and working memory beyond the weights, giving Flash a smaller hardware footprint.

Locate’s Serving Limits

Locate is text-only and capped at 128 Locate questions per request. Requests that mix Locate with decision questions require separate forward passes for each type, adding latency compared with requests containing only one type.

Perplexity keeps the Locate adapters unmerged because merging them caused regressions on ordinary decision requests; they activate only for Locate requests. The design preserves existing decision behavior, but mixed workloads cannot fully share the same computation.

Evaluation and Production Use

Perplexity’s internal evaluation suite contains more than 200 tasks and runs zero-shot, without task-specific fine-tuning. It supplements public leaderboards with private production tasks and held-out evaluations from public non-benchmark datasets to check generalization beyond familiar evaluation suites.

Across 72 retrieval tasks, the company reports a mean nDCG@10 of 83.99 for zero-shot pointwise reranking, in which each query-document pair is scored independently, placing it 1.81 points above Qwen3-Reranker-8B.

For a production check, Perplexity sampled one million deduplicated API requests with personal information masked from the API’s first 18 days of availability. The sample covered 2,318,887 individual questions, all included with explicit user permission for research. Frontier LLM judges supplied reference answers because production requests lack ground-truth labels.

Perplexity reports 93.98% agreement with those judges for v1.3, against 93.68% for v1.2. The difference is modest, and the metric establishes agreement with model-generated references rather than accuracy against ground-truth labels.

For the same request sample, Perplexity reports costs of $26.47 for Decider v1.3 and $13.24 for Flash. Decider ran locally; the comparison does not specify local hardware-cost accounting or provide matched latency results, so the figures describe this workload rather than a universal cost or speed advantage.

Perplexity also uses Decider offline to review agent security logs for missed problems and false alarms. In one reported day, it processed 1 billion tokens across 7,000 agent sessions and identified flaws leading to 14 proposed rule fixes, six of which were approved and shipped within two days. Replaying a week of logs with the fixes cut the busiest rule’s volume by 52%, while widened rules caught about 40% more legitimate actions to investigate.

A second internal system stores probabilities about reinforcement-learning failures and agent interpretations, allowing thresholds to change without new Decider queries. It only observes and diagnoses, never altering rewards or stopping runs. Decider also filters low-quality crawled pages while assessing audience, domain, harmful or illegal content, and personal information in the same pass.

Perplexity recommends calibrating decision thresholds on representative labeled examples because label frequencies and the costs of false positives and false negatives vary by application. It positions the full model for nuanced judgments where errors are costly and Flash for high-volume work constrained by latency and cost.

Decider’s breadth therefore comes with a deployment burden: the same probabilities that make it reusable across tasks still require application-specific thresholds, and Locate’s added extraction capability brings extra computation to mixed requests.