Perplexity releases pplx-decider, an open decision model built on Qwen

Perplexity has released its decision model pplx-decider-v1-27b on Hugging Face. The model is fine-tuned from Qwen3.8-27B and published under the Apache 2.0 license, so anyone can download the model files and run it on their own servers. According to Perplexity's developer announcement, the same model powers its Decisions API, which opened in October 2026.

A decision model is built for the moments inside a workflow where software or an AI agent needs to make a quick call. You give it some content and questions with predefined answers, and it returns a probability for each option without writing a long reply. An application can use those numbers to route a support ticket to the right team, decide whether a message is urgent, or hand an uncertain case to a person. Cloudflare's Clef models, which GlobalFeed covered earlier, belong to the same category.

The model answers three kinds of questions

According to Perplexity's Decisions API documentation, the model supports three question types. For yes/no questions it returns the probability of "yes". For choice questions it assigns a probability to each option and returns the most likely option along with a confidence value. For score questions it rates the content on an ordered rubric, returning a probability for each level and an expected score calculated from those probabilities. A single request can ask several questions about the same content. Inputs can be text, JSON or images.

The examples in the documentation show where the model fits: triaging support tickets, picking the next step for a browser agent, and judging whether the citations in an answer actually support its claims. The code sample on the model card asks whether a customer whose Stripe integration keeps failing should go to billing, technical support or sales.

Close results with Jev in Perplexity's comparison

The model card reports accuracy on 11 benchmarks. In the table Perplexity published, pplx-decider leads overall with 85.71 percent, ahead of Typesafe AI's decision model Jev at 84.51 percent and the base Qwen3.8-27B at 74.76 percent. The widest gap appears on RAGTruth, which measures how well hallucinated content in AI answers is caught. Jev leads on six tests, including WinoGrande, BBH and TruthfulQA. According to the model card, the pplx-decider results were measured through the Perplexity API.

Running it yourself or through the API

To run the model locally, the model card calls for Python 3.12 or later and a CUDA GPU with room for roughly 49 GiB of model files plus working memory. The Hugging Face repository includes an example script that runs with the uv tool and a Python class called Decider for direct use; image inputs can be tried with the same script. The hosted version is available through Perplexity's Decisions API, and the documentation also lists a v1.1 version of the model. Current prices are on Perplexity's site.

The model card does not describe the training data, limitations, or which languages the model handles and how well, so it is worth testing it on your own examples before using it on non-English content. Running it locally requires a data center class GPU. For a fuller picture, follow evaluations from independent researchers as well.

Sources: Hugging Face: pplx-decider-v1-27b model card, Perplexity: Decisions API documentation, Perplexity: developer changelog.