AI Search

Perplexity Ships Open-Source Decider v1.3 Decision Models

Perplexity has released Decider v1.3 and a smaller distilled Decider Flash v1.3, both published under the Apache licence on Hugging Face. The larger model scores 91.6% mean accuracy across 63 query-classification tasks — and it is the kind of small, fast, cheap model that quietly powers the routing decisions inside every AI search product.

What Decider is for

Decider is a decision model, not a chatbot. Its job is to look at a query and decide what to do with it: which category it belongs to, which system should handle it, whether it needs a web search, a calculation, or a straight answer. Every AI search product makes hundreds of these routing decisions behind the scenes, and getting them wrong is how a simple question ends up with a slow, expensive answer — or a hard question gets a cheap one.

The larger model is built on Qwen3.8-27B, an open-weight base, and both models are published under the Apache licence on Hugging Face. That matters because it means anyone can download, run and build on them without a licence fee or a usage agreement. Perplexity is giving away the component that makes the capability work.

Decider v1.3 scores 91.6% mean accuracy across 63 query-classification tasks, which Perplexity says beats its own production classifiers by 6.54 percentage points. That is the company's own comparison, so it should be read as a claim rather than an independent benchmark — but the direction is what matters. A small model that classifies queries better than the production systems it replaces is a genuine improvement, and it is the sort of improvement that compounds across every query a product handles.

The new locate capability

The headline addition in v1.3 is a new locate capability, which extracts verbatim text spans with character offsets and confidence scores. In plain terms: given a document and a question, locate finds the exact passage that answers it and tells you precisely where it sits in the text, with a score for how confident it is. That is the building block for grounded answers — the difference between an AI that says something plausible and one that points at the sentence it is relying on.

Locate adds named-entity recognition, PII detection and extractive question answering to the same API contract, so one small model can do several jobs that used to require several models or calls. For developers building search or retrieval products, that consolidation means fewer moving parts and API calls. The trade-off: locate is text-only and capped at 128 questions per request, so it is a tool for focused extraction, not bulk processing.

The use cases are easy to picture: an SEO tool that pulls the exact passage a page uses to answer a question, a research product that verifies citations against source text, a compliance check that finds personal data in a document. All of those are jobs where the answer has to be verbatim and verifiable, which is exactly what locate is built for.

Pricing that undercuts the big models

The economics are the other half of the story. Running a million production requests costs $26.47 on Decider v1.3 and $13.24 on Flash; in the Decisions API, v1.3 costs $0.02 per million input tokens and Flash $0.01. Those numbers are tiny next to sending the same decisions to a large general-purpose model — routing a query through a frontier model is like hiring a senior lawyer to sort your mail, and a small specialist does it faster for a fraction of the price.

Both models were free on OpenRouter until November 19, 2026, per a Perplexity post on X — a classic distribution play: let developers try the models at no cost, get them into real products, and let the pricing do the rest once the trial ends. For anyone building an AI search or retrieval tool, the free period was the cheapest possible time to test whether Decider fits.

The release also fits a pattern: Perplexity keeps publishing components of its stack, making it easier for smaller teams to build search products that behave like Perplexity's. That is good for the ecosystem and for Perplexity's position in it — the more products built on its models, the more its approach becomes the default.

Why it matters for SEOs and link builders

For SEOs, Decider is a reminder of how the systems that judge your content actually work. The models that decide whether a query needs a web search, which sources to trust and which passage to cite are small, fast classifiers like this one — not the giant chat models that get all the attention. Clear, well-structured content wins because a classifier categorises and locates answers more easily in a page that states them plainly.

The locate capability has a direct SEO implication too. If more tools start extracting verbatim passages to verify citations, then the exact wording on your page matters more than ever. A page that states a fact in one clean, quotable sentence is easier to cite accurately than a page that buries the same fact in a paragraph of hedging. Write for the extractor, not just the reader.

For link builders, the open-source angle is the opportunity. Tools built on Decider will need content, benchmarks and guides — and the teams building them will link to the best resources they find. A clear, practical explainer of how decision models work in search, or a benchmark of how different models classify SEO queries, is exactly the kind of content this release creates demand for.

Sources

Sources: Perplexity announcement, November 12, 2026.

Related reading on this site: YouTube Takes 24.5% of AI Overview Citations in Ahrefs' November Ranking, Wikipedia Leads ChatGPT Citations at 10.1% in Ahrefs' November Ranking, Direct Answers Win AI Citations: 193,507 Answers Studied, Google: Technical SEO Remains Key for AI SEO, Gemini 4 Hydrogen Is Now the Default Model in AI Mode for Everyone, and How Often Sites Block AI Crawlers: 14,318 Sites Measured.

More from Linkbaits

Turn this playbook into links

Linkbaits helps you create content that earns citations and backlinks. Start with the toolkit and build your first linkable asset this week.

Start free trial