Perplexity has released Decider v1.3 and Decider Flash v1.3, the latest versions of its open-weight artificial intelligence decision models, introducing a new capability called “Locate” for extracting specific information directly from text. Announced on November 12, 2026, the models are available under the Apache 2.0 licence, allowing developers to download their weights, customise them and run them on their own infrastructure. Perplexity is also making both models available free through OpenRouter for one week.
Unlike conventional AI chatbots that generate natural-language responses, the Decider models are designed to make structured decisions. They can classify information, evaluate multiple choices, assign probabilities and identify specific text passages. With the addition of Locate, Perplexity is expanding the models’ capabilities into named-entity recognition, personal information detection and evidence extraction, potentially making them useful for developers building AI agents and automated business workflows.
What Are Perplexity Decider v1.3 and Decider Flash v1.3?
Decider is a family of AI models developed by Perplexity for structured decision-making rather than general-purpose conversation.
Instead of generating paragraphs of text, the models analyse an input and return structured outputs, including probabilities for different answers. Developers can provide a piece of text, structured data or an image and ask one or more questions about it.
The models support three main decision formats:
- Yes/no decisions: Estimate the probability that a particular statement or condition is true.
- Multiple-choice decisions: Assign probabilities to different options and identify the most likely answer.
- Scoring decisions: Evaluate information against an ordered scoring scale and calculate an expected score.
For example, a company could provide a customer support conversation and ask the model whether the customer is requesting a refund, reporting a technical problem or asking for product information.
Rather than generating a lengthy explanation, the model returns structured probabilities that software can use to determine the next action.
This approach is particularly useful for applications where predictable outputs and machine-readable results are more important than conversational responses.
New Locate Feature Extracts Exact Text From Documents
The biggest addition in Decider v1.3 is Locate, a feature that allows developers to describe what information they want to find and receive the matching text directly from the supplied content.
For example, a developer could provide a contract and ask the model to locate the name of the person who signed it. The model can return the matching text along with its starting and ending character positions.
The output also includes a confidence score, helping applications assess how reliable an extraction result may be.
Unlike traditional extraction systems that often require developers to create separate rules for different document formats, Locate allows users to describe the information they want in natural language.
How Locate works
The feature can support several types of information extraction:
| Use case | Example |
|---|---|
| Named-entity recognition | Extract company names, people and locations |
| Personal information detection | Locate email addresses and other personal details |
| Contract analysis | Find the person who signed an agreement |
| Document processing | Extract specific dates, amounts and clauses |
| Extractive question answering | Return the exact sentence or passage answering a question |
Locate returns the original matching text rather than generating a rewritten answer. This distinction is important for applications that need to preserve evidence and trace an extracted result back to its original source.
Perplexity says the feature is designed for batch-mode decision-making and evidence or entity extraction. It can be used alongside the models’ classification capabilities, allowing developers to identify information and make decisions about it within the same broader workflow.
The Locate feature operates on text input, even though the underlying Decider models also support image-based decision tasks.
Decider v1.3 vs Decider Flash v1.3
Perplexity has released two versions to serve different computing and performance requirements.
Decider v1.3 is the larger model, while Decider Flash v1.3 is a smaller version designed to offer a more lightweight alternative.
The Flash model is intended for developers who need structured decision-making at higher volumes or want to deploy the technology with fewer computing resources.
| Feature | Decider v1.3 | Decider Flash v1.3 |
|---|---|---|
| Model size | Approximately 26B parameters | Approximately 5B parameters |
| Model type | Decision model | Smaller decision model |
| Locate feature | Yes | Yes |
| Input support | Text and images | Text and images |
| Output | Structured decisions and probabilities | Structured decisions and probabilities |
| Open-source licence | Apache 2.0 | Apache 2.0 |
| Primary use | Complex classification and extraction | Lightweight, high-volume workflows |
Both models support the same general decision-making interface. Flash is a smaller alternative that developers can consider when computational requirements and deployment costs are important.
Perplexity Reports Improved Accuracy Across Classification Tasks
Perplexity says the updated models deliver improvements across its internal classification workloads and several external evaluation tasks.
According to the company’s reported results, Decider v1.3 achieved 91.6% mean accuracy across 63 query-classification tasks. That represents a 6.54 percentage-point improvement over the production classifiers Perplexity was previously using.
The company also evaluated the models across 17 out-of-domain tasks involving named-entity recognition, personal information detection and extractive question answering.
For the Locate capability, Perplexity reported average scores of 69.0 for Decider v1.3 and 63.6 for Decider Flash v1.3.
Other reported results include:
| Evaluation | Decider v1.3 result |
|---|---|
| Mean accuracy across 63 classification tasks | 91.6% |
| Improvement over existing production classifiers | 6.54 percentage points |
| Locate average score across 17 tasks | 69.0 |
| Multilingual evaluation | 88.16% |
| Image classification | 76.99% |
| Agreement with frontier AI judges on production requests | 93.98% |
These figures come from Perplexity’s own evaluations and should not be treated as independently verified benchmarks. Performance can also vary depending on the type of information being processed and the application in which the models are deployed.
Models Available Free on OpenRouter for One Week
Perplexity is making Decider v1.3 and Decider Flash v1.3 available free through OpenRouter for one week following the November 12 launch announcement.
OpenRouter provides a unified interface that allows developers to access AI models from different providers through a common API platform.
The free listings make it easier for developers to experiment with the models without immediately paying for API usage. However, free endpoints may have rate limits and availability restrictions.
Developers can access the models through these listings:
The models are also available through Perplexity’s Decisions API and on Hugging Face.
The free OpenRouter offer is separate from the open-weight release. Developers can continue using the downloadable model weights under the Apache 2.0 licence, subject to the licence terms, even after the promotional API period ends.
Developers Can Download and Run the Models Locally
Open-sourcing the model weights gives developers more deployment flexibility.
Instead of depending entirely on a hosted API, organisations can download the models and run them on compatible infrastructure. This could be useful for businesses that want greater control over their data processing, model deployment and operating costs.
The larger Decider v1.3 requires substantially more computing resources than Flash. The smaller model is therefore a more practical option for some local deployment environments.
The Apache 2.0 licence also allows commercial use, modification and redistribution subject to its conditions.
For developers, the combination of open weights, structured decision outputs and Locate could make the models useful for building document-processing systems, AI agent safeguards, automated moderation tools and enterprise information extraction pipelines.
The Bigger Picture
Perplexity’s latest release highlights a growing distinction in the AI industry between models designed to generate content and models built to make structured decisions. While general-purpose AI assistants focus on writing, reasoning and conversation, decision models can perform specialised tasks such as classification, ranking, routing and information extraction.
The introduction of Locate makes this approach more practical by combining decision-making with the ability to identify exact evidence within source material. This could help developers build AI systems that not only determine what action should be taken but also identify the information supporting that decision.
The open-weight release is equally significant. By publishing the models under Apache 2.0 and offering temporary free access through OpenRouter, Perplexity is giving developers multiple ways to test and deploy its technology. However, the company’s benchmark claims will need to be evaluated against independent testing and real-world workloads before broader conclusions can be drawn about their competitiveness.
Looking Ahead
The one-week OpenRouter promotion gives developers an opportunity to test Decider v1.3 and Decider Flash v1.3 without paying for API usage during the offer period. Their adoption will depend on how well the models perform in practical classification, extraction and automated decision-making applications. Developers will also need to assess the computing requirements of self-hosting the larger model against the lighter Flash alternative.
For Perplexity, the release expands its presence beyond consumer-facing AI search and conversational products into the developer infrastructure market. As businesses increasingly build AI agents that need reliable classification and evidence extraction, specialised decision models could become an important component of automated software systems. The new Locate capability gives Perplexity another way to compete in that emerging category.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
