Perplexity releases Decider 1.3 under Apache on Hugging Face

A team pays a frontier model to answer the same yes-or-no question a million times a month. The bill is large and the answer is a single token. Perplexity has opened a smaller model trained for that job.

Perplexity released Decider v1.3 and a smaller Flash version. Both are on Hugging Face under the Apache licence. The company also offers the model through its Decisions API and through OpenRouter. The post names Antoine Chaffin, Jonas Golde, Louis Milliken, Sarah Eslami and others as authors.

What a decision model is

A decision model sits between a full language model and a task-specific classifier. It takes an input and returns probability scores for a set of outcomes. It never writes text. Perplexity says it uses Decider across 63 internal tasks. Those tasks include query classification, web page quality filtering, and monitoring reinforcement learning rollouts. The model also finds gaps in security rules.

What is new in 1.3

The company names four changes. First, bidirectional attention lets each token read the whole input. Second, the training data is better curated and the training method is refined. Third, a new Locate feature can pull text from the input instead of picking from preset options. The feature covers named entity recognition and information extraction. Fourth, the Flash variant is smaller, distilled from the full model.

The published figures

Perplexity reports three benchmark groups, all its own.

On query classification, the company says Decider v1.3 reaches 91.6 percent mean accuracy. That is 6.54 percentage points above its own production classifiers. On 2.3 million real API requests, Perplexity reports 93.98 percent agreement between Decider and frontier language model judges. On retrieval and reranking, Decider scores 83.99 nDCG@10 across 72 tasks. The company says the model leads across seven domains. In finance it beats its top competitor by up to 12.9 points.

The cost claim

Perplexity reports a cost comparison for one million production requests. Decider v1.3 costs 26.47 US dollars. Perplexity names competitor prices between 68 and 349 US dollars for the same workload. The figure is Perplexity's own, from its own API pricing. A buyer should test the same workload before trusting the comparison.

What a team should do with it

Decider is for teams that need a probability score, not a sentence. A support router, a product tagger, and a safety filter all fit the shape. A web crawler quality check and a retrieval reranker fit too. A team that already pays a frontier model for a yes-or-no question should test Decider on its own data. The Apache licence lets a team run the weights on its own hardware. The Hugging Face collection carries both the full model and the Flash variant.

Source

• Perplexity Decider v1.3, Perplexity blog.
• Decider v1.3 collection, Hugging Face.