Perplexity has released Decider v1.3 and Decider Flash v1.3, its open-weight decision models, under the Apache license — and added “locate,” a feature that returns the exact text spans matching a request instead of just a probability. Why it matters: a growing share of AI work is not writing prose but making fast, cheap judgments — classify this, route that, flag this — and decision models do that work in a single pass, without generating text token by token.
November 13, 2026
What happened
Perplexity announced the two models on November 12. Both are available as downloadable weights on Hugging Face under the Apache license, through OpenRouter, and through the Perplexity Decisions API. Decider v1.3 is built on a Qwen3.8-27B backbone; Decider Flash v1.3 is a smaller Qwen3.5-4B model distilled from it. Per its API changelog, pricing is unchanged from v1.2: $0.02 per million input tokens for Decider v1.3 and $0.01 for Flash, with output tokens free. Both accept text and images and can answer several questions about the same input in one pass.
What is new in v1.3
The headline addition is locate. A locate question describes what to find — a short label such as “location,” or an instruction such as extracting the city name in a job posting — and the model returns the exact matching text with its start and end character positions, or an empty list when nothing matches. Perplexity built it for named-entity recognition, personal-information detection and extractive question answering. Per the changelog, locate works on text only and does not work with web grounding. The v1.3 models also switch to an “isolated questions” attention pattern, which Perplexity says makes each answer more stable when several questions share one input, and use bidirectional attention in the backbone’s full-attention layers.
What the evaluations show
All figures here are Perplexity’s own. Across 63 classification tasks with references independently relabeled by a frontier LLM, Decider v1.3 reached 91.6% mean accuracy, 6.54 percentage points above Perplexity’s production classifiers. On a sample of one million real Decisions API requests, v1.3 agreed with frontier-LLM judges 93.98% of the time, up from 93.68% for v1.2. On 17 out-of-domain extraction tasks, Perplexity reports an average score of 69.0 for Decider v1.3 and 63.6 for Flash, against 43.3 for Liquid AI’s d1. Perplexity also says it uses Decider internally for security-rule review, RL rollout monitoring and page-quality filtering while crawling.
Why it matters
Decision models are becoming a distinct layer of the AI stack, alongside generative models: Google, Cloudflare, Microsoft and others have all shipped one this year. Perplexity’s entry is notable because the weights are open, the API price is pennies per million tokens, and locate turns the same model into an extraction tool — the kind of unglamorous, high-volume work (moderation queues, routing, PII scans, reranking) where most production AI spend actually goes. If the reported numbers hold up in outside testing, a 4B open model that classifies and extracts at this level is hard to argue with on cost.
Frequently asked questions
What is a decision model?
A general-purpose classifier: you give it an input and a set of questions, and it returns a probability distribution over the possible answers for each one, in a single forward pass, instead of generating text.
What does Decider v1.3 cost?
Through Perplexity’s Decisions API, $0.02 per million input tokens; Decider Flash v1.3 costs $0.01 per million input tokens, and output tokens are free. The weights are also free to download under the Apache license.
What is the new locate feature?
Locate returns the exact spans of text that match a described field — with character offsets — for tasks such as named-entity recognition, PII detection and extractive question answering. It is text-only and does not work with web grounding.
Sources: Perplexity (Decider v1.3 announcement and Decisions API changelog), Unite.AI