Skip to main content

Perplexity releases Decider v1.3 decision models

13 NOVEMBER 2026·2 MIN READ·2 SOURCES

Perplexity announced Decider v1.3 and Decider Flash v1.3 on Nov. 12, 2026, describing them as successors to its earlier decision models. The company says both are available on Hugging Face under an Apache license, through OpenRouter and through its Decisions API.

Perplexity releases Decider v1.3 decision models

Key takeaways · 4

  • 01

    Both models are available through Hugging Face, OpenRouter and the Decisions API; Perplexity lists an Apache license for the Hugging Face versions.

  • 02

    Decider supports binary and multi-class classification and returns probabilities; Perplexity says applications can use it to classify, route, rank and evaluate inputs.

  • 03

    Flash is the smaller model, distilled from Decider v1.3; Perplexity lists its price at $0.01, compared with $0.02 for Decider v1.3.

  • 04

    Perplexity reports 91.6% mean accuracy on its internal suite and a mean nDCG@10 of 83.99 across 72 retrieval tasks.

Availability and core capabilities

Perplexity announced Decider v1.3 and Decider Flash v1.3 as successors to its earlier decision models.[1] The company says both models are available on Hugging Face under an Apache license, through OpenRouter and through its Decisions API.[1] Decider supports binary and multi-class classification, with ordered or unordered labels, and returns probabilities.[1] Perplexity says applications can use it to classify, route, rank and evaluate inputs.[1] That combination positions the models for workflows that need a decision or score from an input rather than a free-form response.

Locate handles extraction tasks

Perplexity describes a Locate capability for named-entity recognition, personal-information identification and extractive question answering.[1] Its output includes matching text and the start and end character locations; when there are no matches, it returns an empty list.[1] Perplexity says it keeps Locate adapters unmerged and activates them only for Locate requests because merging is not yet lossless.[1] Teams assessing the feature should account for that separate adapter behavior when planning how Locate requests fit into a workflow.

How Perplexity says it trains them

Perplexity says it converts examples from hundreds of public datasets into a shared format containing a state, questions, answer options and a target.[1] Its generalist training stage excludes benchmark training splits, followed by a domain-focused stage using LoRA to limit drift from the generalist model.[1] The company says it averages weights from three separately seeded runs for both text-training stages and removes samples that could overlap with evaluations to mitigate contamination.[1] It also says the smaller model is trained to minimize the difference, measured as KL divergence, between its output distribution and the larger teacher model’s.[1]

Reported results and listed prices

Perplexity reports that Decider v1.3 achieved 91.6% mean accuracy on an internal suite of more than 200 tasks, 6.54 percentage points above its production classifiers.[1] The company says the suite includes classification, scoring, reranking and entity extraction, and that evaluations are zero-shot, without task-specific fine-tuning.[1] It also reports a mean nDCG@10 score of 83.99 across 72 retrieval tasks, 1.81 points above Qwen3-Reranker-8B, and says Decider outperformed Jev 1.14 in all eight languages in its multilingual classification evaluation.[1] Listed prices are $0.02 for Decider v1.3 and $0.01 for Flash; output tokens are free.[2]

For teams evaluating decision models, the announcement provides a choice of model size, access route and listed price, alongside reported results on classification and retrieval tasks. The training and evaluation details are useful context, but Perplexity’s reported results are not a substitute for testing the models on a team’s own tasks.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 13 November 2026

    Perplexity releases Decider v1.3 decision models

Sources

AI fluency, one session a day, built for your work.