AIGenerated
EN · IT · ES
Models

Decider v1.3

Decider v1.3 is an open‑source AI model from Perplexity AI that classifies text or images into predefined labels instead of generating long answers, making it much faster and cheaper than a chatbot for sorting content or extracting specific information.

2 stories, first covered on Nov 14, 2026.

What Decider v1.3 does

Decider v1.3 is an artificial‑intelligence model that categorises text or images rather than writing paragraphs of prose. Instead of producing a long reply, it returns a probability score for each possible label – yes or no, multiple‑choice options, or a rating with a confidence measure. This makes it far quicker and less expensive than a regular chatbot for jobs such as sorting emails, checking whether content is safe, or answering several classification questions in one call.

How it works

The model was built on top of Qwen3.8‑27B, a large generative language model. Perplexity trimmed its output layer down to just the answer choices you want (for example A, B, C) and kept most of the backbone frozen. It also added small fine‑tuning adapters and trains using bidirectional attention, so every word can look at both earlier and later words in the same input. This shared processing lets the model answer many questions at once while “sharing computation across every question.”

A new feature called Locate goes beyond yes‑or‑no: it pulls out exact snippets of text together with their start and end character positions. If you ask it to find all city names in a document, it returns each city along with exactly where it appears.

Both Decider v1.3 and a smaller Flash version are open‑source under an Apache licence. They can be downloaded from Hugging Face or used through online APIs hosted by Perplexity or OpenRouter.

Where it stands now

Perplexity reports the full 27‑billion‑parameter model scored 91.6 percent mean accuracy across 63 internal query‑classification tasks, beating TypeSafe’s Jev 1.14. On retrieval rankings, multilingual classification and image scoring it led; Locate averaged 69.0, lower than the other categories. For real‑world workloads Perplexity recorded 93.98 percent agreement with frontier LLM judges on one million production requests, compared with 93.35 percent for Jev 1.14. Answer flips caused by changing the order of questions fell to 0.27 percent versus up to 4.84 percent for Jev.

Through the API the cost is $0.02 per million input tokens for the full model; pricing for the Flash variant was not disclosed.

Questions people ask

Is Decider v1.3 free to use?

Yes, both Decider v1.3 and the smaller Flash version are open‑source under an Apache licence, meaning anyone can download and run them on their own machines or access them through third‑party APIs. Perplexity also offers an API, but using it costs money based on how many tokens you send.

What kind of tasks is it good for?

It is built for classification jobs where you need to sort, label, or extract specific pieces of information from text or images – for example, flagging spam, judging whether an AI summary stays faithful to a source, routing difficult requests to a bigger model, or pulling out exact quotes with their positions.

How does it compare to a regular chatbot?

Unlike a chatbot that writes long answers word by word, Decider v1.3 quickly returns probability scores for predefined labels. This makes it much cheaper and faster for high‑volume filtering or decision‑making, though it cannot generate creative text or explanations.

Written by AI only from the stories listed below, each of which links to its own sources.

Every story about Decider v1.3 (2)