Perplexity has announced Decider v1.3 and Decider Flash v1.3, adding a feature called “locate” that finds text based on instructions. In his November 12, 2026, announcement, Arav Srinivas said the feature returns matching text with character offsets, and described Perplexity as open-sourcing both models. Source
The pitch is a highlighter that takes instructions, rather than another invitation to write an essay. Those character offsets identify where the selected text sits in the input, giving developers a way to connect an extracted snippet back to its original location. Source
What “locate” is for
A separate launch post from denisyarats lists intended uses including named-entity recognition, personal-information detection and extractive question answering. In plain English: picking out names and other entities, locating potentially identifying information, or selecting an answer from supplied text rather than composing a fresh response. Srinivas also points to batch decision-making and evidence or entity extraction. Source Source
A pointer is not a truth stamp: locating a passage doesn’t, by itself, verify that it’s true.
The release also prioritizes robustness, according to denisyarats. On 63 of Perplexity’s own production classification tasks, Srinivas says v1.3 was 6.54 points more accurate than the classifiers the company was already running. His post doesn’t define the metric or explain how the results were aggregated. Source Source
The launch offer has a clock attached
In the November 12 announcement, Srinivas said the weights were open on Hugging Face and the models were being made free on OpenRouter for a week. His post doesn’t specify exact offer times, usage limits or license terms. Source
Comments