Perplexity Open-Sources Decider 1.3, Its Judge for AI Agents

Human Verified 0

Share to: , and

A glowing teal network of branching decision pathways converging on a balance node against a dark navy background

Perplexity released Decider 1.3 and Decider Flash 1.3 on November 12, two open-weight models that make the small decisions inside AI systems. Instead of generating text, they classify, route and score — returning probabilities for each possible answer. Both are free to download under the Apache licence.

What a decision model does

A decision model is a classifier, not a chatbot. Give it a document and questions, and it returns a probability for each possible answer in one pass. It can judge whether a summary is faithful, route work to a cheap or powerful model, or flag harmful content.

That matters because most AI work is not free-form writing. Agents constantly make small calls — is this page useful, is this tool output safe, which model should answer — and doing those with a full language model wastes time and money. Decider answers with numbers, not prose.

What is new in 1.3

The headline addition is locate, a feature that pulls matching text out of a document and returns its exact position. Ask for a person’s name, an email address or the total on an invoice, and the model returns the span itself — not just a yes or no.

Under the hood, the larger model is built on Qwen3.8-27B, while Flash is a smaller distilled version on Qwen3.5-4B. Perplexity rebuilt the attention so each question is answered in isolation, which keeps results stable when several questions share one document.

On Perplexity’s own tests, Decider 1.3 averaged 91.6% accuracy across 63 classification tasks, beating the company’s production classifiers by 6.54 points. It also led across seven domains — finance by 12.9 points and law by 9.3 — against Jev 1.14, the rival decision model it is modelled on.

How Perplexity uses it

Perplexity already runs Decider inside its own systems. Its security team uses it to audit agent sessions: in one day it read a billion tokens across 7,000 agent sessions and found flaws that led to 14 rule fixes, six of which shipped within two days.

The model also watches reinforcement-learning rollouts, diagnosing why an attempt failed — a broken tool, a bad sandbox, a made-up claim — without changing any reward. And it filters web pages during crawling, judging quality, audience and harmful content in a single pass.

Our read

Decision models are the plumbing of the agent era, and Perplexity is betting that plumbing should be open. By publishing the weights under Apache, it invites developers to run the judge locally, tune thresholds on their own data, and stop paying a frontier model to answer yes or no.

The numbers deserve a careful read. Every score in the release is Perplexity’s own, and the company warns that leaderboards can be gamed by repeated submission. The independent check so far is Artificial Analysis, which placed Decider 1.3 sixth on its Decision Index with a score of 70.

Still, the design choices are the interesting part. Isolated questions, a locate head kept separate so it does not degrade ordinary decisions, and a cost of two cents per million input tokens all point at a model built to be used constantly, in the background, unnoticed.

Why it matters

Agents are only as good as the small decisions they make between the big ones. A cheap, open, reliable judge lowers the cost of every agent workflow — routing, moderation, evaluation — and gives smaller teams the same plumbing the big labs run.

Background

Decider is the successor to Perplexity’s earlier decision models, versions 1 through 1.2, which the company began shipping in October. The Decisions API serves the models at two cents per million input tokens for the full model and one cent for Flash, with output tokens free.

Source: Perplexity blog post and AlphaSignal report

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.