What happened
Perplexity’s announcement says Decider v1.3 and Decider Flash v1.3 succeed its earlier decision models. The full model is based on Qwen3.8-27B, while Flash is a smaller model distilled from it and based on Qwen3.5-4B. Both models are available on Hugging Face under the Apache license, through OpenRouter, and through Perplexity’s Decisions API.
Perplexity describes Decider as a general-purpose classifier. An application supplies a shared input state and one or more questions with possible answers; the model returns a probability distribution over those answers. It supports yes/no decisions, custom multiple-choice labels, and ordered score levels. The company positions these outputs for judging model responses, routing requests by difficulty, moderation, prompt-injection detection in tool outputs, content tagging, retrieval, and reranking. The intended distinction is that these jobs need bounded decisions rather than generated prose.
The new capability is called locate. Perplexity says it extracts matching text from the input state and returns the matching text with start and end character positions. The feature is intended for named-entity recognition, personal-information identification, and extractive question answering. A request can combine locate, classification, and scoring over the same shared input. Perplexity says locate is trained on span-extraction tasks with field descriptions ranging from short labels to detailed instructions, and returns an empty list when it identifies no match.
The announcement also details changes to the model architecture. Perplexity says v1.3 uses bidirectional attention and an “isolated questions” pattern: the shared state attends to all questions, while each question attends only to the state and its own tokens. The company chose this design for performance and answer stability. It describes a fully cacheable alternative in which state processing is independent of questions, but says that mode is not the default because allowing the state to see the questions matters substantially for some tasks. Perplexity also says the locate adapters remain unmerged because merging them into the backbone currently causes regression on ordinary decision requests.
Perplexity reports results from internal and external evaluations. Across 63 query-classification tasks, it says Decider v1.3 achieved 91.6% mean accuracy, 6.54 percentage points above its production classifiers. Across 72 retrieval tasks, it reports a mean nDCG@10 of 83.99, 1.81 points above Qwen3-Reranker-8B. On 17 out-of-domain extraction tasks, Perplexity reports average scores of 69.0 for v1.3 and 63.6 for Flash, versus 43.3 for Liquid d1, 41.5 for Clairic, and 33.6 for GLiNER large v2.5. The company says these extraction tasks use their respective primary metrics, including exact-span F1, character-level F1, and field accuracy.
What the archive adds
New Runtime’s earlier coverage includes Decider for Mac running Flash v1.2 locally on 2026-10-24 and the v1.2 decision-model release on 2026-11-02. This is a new version on a known story: it adds span extraction and revised evaluation claims to the existing classification and scoring product line.
Trend signal
This release supports model answers becoming task interfaces. Decider returns bounded probability distributions over application-defined options and, with locate, selected spans with character positions. That output is designed to feed routing, ranking, review, and extraction workflows rather than a conversational interface. The announcement describes intended uses, not independently measured production reliability across those uses.
Why now
Perplexity presents v1.3 as a successor to several previous Decider releases. The update expands the family from classification and scoring into extraction while changing its attention pattern to improve stability. Perplexity frames the release as a combined capability and robustness update rather than a new product category.
Why it matters
For builders, the operational choice is often between asking a generative model to emit a formatted answer and using a component whose output is already a score over constrained options. Probability distributions can enter routing, ranking, moderation, or review logic directly. Locate adds a way to return extractive results with positions in the original text, potentially reducing downstream parsing and making selections easier to inspect.
The release offers both hosted and self-managed evaluation paths: the models are available through Perplexity’s API and OpenRouter, while Apache-licensed weights are on Hugging Face. That does not remove validation work. A false classification, an incorrect threshold, or an erroneous span boundary can still trigger a bad downstream action. Teams should treat Decider as a candidate component to compare against an existing classifier, reranker, or extraction pipeline on their own labels, documents, languages, latency requirements, and error costs.
Limits and uncertainty
Perplexity reports the evaluation figures and describes its internal suite of more than 200 tasks. Those results have not been independently checked here. The headline averages span different task sets; for locate, the aggregate uses primary metrics that vary by task. They therefore do not establish performance on a particular organisation’s documents, label definitions, risk thresholds, or production traffic.
Perplexity also says that merging locate adapters currently regresses ordinary decision requests and that the adapters are activated only for locate requests. The practical consequences for latency, deployment complexity, concurrent use of capabilities, and cost depend on a team’s chosen serving path and need direct testing.
Skill to build now
Build a small labelled evaluation set for one bounded job this week: route support requests, rank retrieved documents, or extract fields from contracts. Compare Decider with the current system on accuracy, false positives, missed spans, latency, and failure cases with no valid answer. For locate, verify that returned character positions map correctly to the original document after any preprocessing.
Reader takeaway
Decider v1.3 adds extractive span selection to Perplexity’s specialized decision-model family while retaining structured classification and scoring. Its meaningful promise is a defined output interface with available weights and hosted access. Benchmark that interface on a real bounded workflow before using its published averages to justify automation.
Open question
How does locate perform on a team’s own documents, languages, and error-sensitive extraction jobs when its span boundaries drive automated actions?