Perplexity’s Decider v1.3 scores decisions and extracts text

By James Calloway

Perplexity released Decider v1.3 and Decider Flash v1.3 on November 12, 2026, reporting 91.6% mean accuracy for the larger model across 63 query-classification tasks. Perplexity says a frontier large language model (LLM) independently relabeled the references, and the score exceeded its production classifiers by 6.54 percentage points.

Decider v1.3 and Decider Flash v1.3 return probabilities for yes/no questions, custom choices or ordered scores instead of generating prose. Perplexity says these workflows usually need a decision threshold, the cutoff that turns a probability into an action, rather than free-form generation.

An application can ask several questions about the same text, JSON or images, then use the resulting numbers for:

• Classification
• Request routing
• Moderation
• Filtering
• Output evaluation
• Reranking

What changed in the release?

Perplexity made the models’ weights downloadable under Apache-2.0, a permissive software license. Both models are also available through OpenRouter and Perplexity’s Decisions API, widening access beyond a hosted-only release.

Technically, Decider adapts generative backbones by replacing next-token output with a decision head at a marker token. The decision head starts from only the base model’s rows tied to answer-option tokens, converting language-model representations into fixed-choice scores.

In v1.3, Perplexity makes full-attention layers bidirectional so tokens can use context on both sides. The company isolates questions from direct attention to one another while letting their shared input state attend to every question, replacing v1.2 behavior in which one answer could change depending on what else was asked.

Perplexity says the arrangement should make answers more stable. It does not publish prompt-sensitivity results in the release material, so stability remains a design claim rather than a fully documented outcome.

Takeaway: The release combines downloadable scoring models with a specific attempt to stop one question from altering another.

What does Locate add?

Locate returns verbatim matching text with start and end character positions. Perplexity lists named-entity recognition, which labels entities in text, personal-information detection and extractive question answering as intended uses, while a request with no match returns an empty list.

Locate is text-only and capped at 128 Locate questions per request. Requests mixing Locate with decision questions need separate forward passes, or model computations, for each type, which Perplexity says adds latency.

Perplexity keeps Locate adapters unmerged because merging them caused regressions in ordinary decision requests. They activate only for Locate traffic, preserving existing decision behavior but preventing mixed workloads from fully sharing computation.

Takeaway: Locate adds exact text positions, but its separate path introduces computational and latency costs.

How did Perplexity train the models?

Perplexity converts hundreds of public datasets into a shared state, question, options and target format. Generalist fine-tuning excludes benchmark training splits, and the company says it also removes samples that could overlap its evaluations.

Later domain and vision specialization uses LoRA, an adapter method that freezes pretrained weights while training injected rank-decomposition matrices. Perplexity uses weighted adapter merges to control drift, or unwanted changes, and averages weights from three runs for the text stages.

Decider Flash v1.3 is distilled from the larger model. Its training uses KL divergence, a measure of difference between probability distributions, to reduce disagreement with the teacher’s output probabilities 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 in 16-bit floating-point precision. The Decider Flash v1.3 card specifies Qwen3.5-4B and roughly 8.5 GiB of BF16 weights.

Takeaway: The training pipeline prioritizes reusable probabilities and evaluation hygiene, but it relies on several company-controlled filtering and weighting choices.

What serving limits apply?

Self-hosting either model requires a CUDA-capable graphics processing unit (GPU) and working memory beyond the stored weights. Perplexity says this gives Flash the smaller hardware footprint, although its model file does not capture all runtime memory needs.

Perplexity does not publish minimum GPU memory, power draw or matched serving throughput for either model card. Those omissions make the stated file sizes useful for scale, but incomplete for procurement.

Takeaway: Flash needs less stated hardware, but missing memory and throughput figures complicate actual deployment.

How strong are the evaluations?

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

Across 72 retrieval tasks, Perplexity reports a mean nDCG@10 of 83.99 for zero-shot pointwise reranking. The normalized discounted cumulative gain metric evaluates the top 10 results, while pointwise reranking scores each query-document pair independently. Perplexity says the score stands 1.81 points above Qwen3-Reranker-8B.

The 91.6% classification result and 83.99 retrieval result both rely on Perplexity’s own reporting. The company does not identify the 63 query-classification tasks, the production classifiers used as its baseline, or independent replication for either score.

Takeaway: The numbers show a broad internal test program, but missing dataset and baseline details limit outside comparison.

What did production tests show?

Perplexity sampled one million deduplicated, or non-duplicate, Decisions API requests from the API’s first 18 days of availability. It masked personal information and counted 2,318,887 individual questions, with all requests included under explicit user permission for research.

Because production requests lack ground-truth labels, meaning verified correct answers, frontier LLM judges supplied reference answers. Perplexity reports 93.98% agreement for v1.3, compared with 93.68% for v1.2, a modest improvement.

That metric measures agreement with model-generated references, not accuracy against ground truth. The modest gap therefore does not establish how often either version produces a substantively correct answer.

For the same request sample, Perplexity reports costs of $26.47 for Decider v1.3 and $13.24 for Decider Flash v1.3. Decider ran locally, but Perplexity’s release material omits local hardware-cost accounting and equal-condition latency comparisons. The totals describe this workload rather than a universal cost or speed advantage.

Takeaway: Production traffic offers useful scale and cost data, but it cannot establish true accuracy without human ground truth.

What does deployment require?

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 found flaws that led to 14 proposed rule fixes.

Six of those fixes were approved and shipped within two days. Perplexity says replaying a week of logs with the changes 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 activity; Perplexity says it never alters rewards or stops runs.

Decider also filters low-quality crawled pages while assessing audience, domain, harmful or illegal content and personal information in the same pass. Perplexity does not publish category-level accuracy for those checks, so combined processing does not reveal each filter’s reliability.

Perplexity recommends calibrating decision thresholds on representative labeled examples because label frequencies and the costs of false positives, or wrong alarms, and false negatives, or missed cases, 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 same probabilities that make Decider reusable across tasks still require application-specific thresholds. Locate’s extra computation in mixed requests means savings depend on traffic mix, hardware and the cost of each wrong decision.

Takeaway: Decider covers many narrow filters, but buyers must supply labeled examples and benchmark their own hardware before trusting it. Matched hardware cost and latency data would be the next useful disclosure, not another company-reported score alone.