Perplexity’s Decider v1.3 Adds Text-Span Extraction

By Olivia Hart

Perplexity says Decider v1.3 reached 91.6% mean accuracy across 63 query-classification tasks, 6.54 percentage points above its production classifiers. The company released Decider v1.3 and Decider Flash v1.3 on November 12, 2026, adding verbatim text-span extraction and downloadable Apache-2.0 weights. Perplexity’s reported benchmark used references independently relabeled by a frontier LLM.

What does the release add?

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

Perplexity’s Decider returns probabilities for yes/no questions, custom choices and ordered scores instead of prose. One application can ask several questions about the same text, JSON or images, then use the outputs for classification, request routing, moderation, filtering, output evaluation and reranking. Perplexity says these tasks often require a threshold rather than free-form generation, making reusable scores and extracted spans the practical alternative to generated answers.

How does v1.3 improve stability?

Perplexity replaces next-token output with a decision head read at a marker token, initialized from base-model rows associated with answer-option tokens. In v1.3, full-attention layers become bidirectional; questions no longer attend directly to one another, although the shared input state attends to them. Perplexity says the design is “designed to make answers more stable,” replacing v1.2’s arrangement, where another question could change an answer.

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 removes samples that could overlap evaluations.

Later domain and vision specialization uses LoRA, freezing 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 then distilled from the larger model with KL divergence, which Perplexity says reduces differences between Flash’s output probabilities and the teacher’s rather than relying on gold labels alone. The design targets more stable answers, though Perplexity still supplies the evidence.

What does self-hosting require?

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 lists Qwen3.5-4B and roughly 8.5 GiB. Perplexity says self-hosting either model requires a CUDA GPU and additional working memory; Flash has the smaller footprint, but both remain GPU deployments.

Why does Locate add latency?

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

Perplexity keeps Locate adapters unmerged because merging caused regressions on ordinary decision requests; they activate only for Locate requests. That preserves existing decision behavior, according to Perplexity, but mixed workloads cannot fully share computation. Locate therefore adds extra passes and a separate serving path.

How strong are the reported results?

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

Across 72 retrieval tasks, Perplexity reports a mean nDCG@10 of 83.99 for zero-shot pointwise reranking, which scores each query-document pair independently. The figure is 1.81 points above Qwen3-Reranker-8B, according to Perplexity.

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 used with explicit user permission for research. Frontier LLM judges supplied references because production requests lacked ground-truth labels.

Perplexity reports 93.98% agreement with those judges for v1.3, versus 93.68% for v1.2. It calls the difference modest and says the metric measures agreement with model-generated references, not accuracy against ground-truth labels.

For the same sample, Perplexity reports costs of $26.47 for Decider v1.3 and $13.24 for Flash. Decider ran locally, but the comparison omits local hardware-cost accounting and matched latency results, so the figures apply to this workload rather than a universal cost or speed advantage. Perplexity shows a modest gain, not a matched cost or latency advantage.

Where is Decider already working?

Perplexity 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 found flaws leading to 14 proposed rule fixes. Six 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, according to Perplexity. These results show internal operating value, not independent proof for outside customers.

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, broadening its internal workload without establishing independent outside-customer results.

What must customers prove before deployment?

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 Decider v1.3 for nuanced judgments where errors are costly and Flash for high-volume work constrained by latency and cost.

The probabilities remain reusable only after application-specific calibration, while Locate adds computation to mixed requests. The practical test is whether Perplexity’s internal gains survive outside its own logs, traffic and evaluation data.