pplx-decider-v1.2

pplx-decider-v1.2

Reads a situation and several questions in one pass, then answers each with a choice, a yes or no, or a score. Built for teams that route, flag or rate incoming requests.

Made by perplexity-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: pplx-decider-v1.2 is a text classification model from Perplexity that reads a situation and several questions in one pass, then answers each one. It follows the v1.1 27B decision model and uses the same Qwen3.8-27B base. It takes text and images.

Who it suits: Teams that need a model to make structured calls about a piece of text, such as routing a message or rating how upset a customer is. Researchers who compare decision benchmarks can use it too.

Industry | Who uses it | How they use it | Example
Customer support | Support engineers | Send each message to the right team and flag blocked customers | A 401 error after a key change goes to the support team
Software | ML engineers | Add a classification step that answers several questions at once | Tag one ticket with a team, an urgency flag and a mood score
Research | Evaluation researchers | Compare decision benchmarks across model versions | Run the same suite with joint and single-question requests

If you have no GPU with about 52 GiB of room, you can skip running it.

Where to use it:

• A Python script or service on a machine with a CUDA GPU
• Linux, Windows or macOS for downloading the files
• Text requests, and images up to 1.6 MP

Use case | Who | What they get back
Request routing | Support teams | A chosen team for each message
Urgency check | Operations staff | A yes or no on whether the customer is blocked
Mood scoring | Support leads | A score picked from a list of levels you supply

How to set it up:

• Python 3.12 or newer
• A CUDA GPU that can hold roughly 52 GiB of BF16 weights and extra memory to work in
• A Hugging Face token with access, since the repository is private

1. Install the Hugging Face hub package.
2. Download the repository to your computer.
3. Install the packages pinned in the requirements file inside the downloaded folder.

pip install -U huggingface_hub
hf download perplexity-ai/pplx-decider-v1.2

The source folder can also be installed as a package:

pip install ./source

How to use it: Load the checkpoint, build a Decider on the GPU, then call decide with a request that holds a state and a set of questions. Each question has a type, such as choice, noul or score.

import sys
from pathlib import Path
from huggingface_hub import snapshot_download

checkpoint = Path(snapshot_download("perplexity-ai/pplx-decider-v1.2"))
sys.path.insert(0, str(checkpoint / "source" / "src"))
from pplx_decider import Decider

decider = Decider.from_pretrained(checkpoint, device="cuda")

Tool | What it does
Decider.from_pretrained | Loads the checkpoint onto the CUDA device
decide | Takes one request and returns one response, or a list of responses for a list of requests

Example: A customer writes that their integration fails with a 401 since they rotated an API key. You ask which team should handle it, whether the customer is blocked, and how frustrated they sound. The response holds one answer per question, such as a team choice and a yes or no on urgency.

For newcomers and experts: If you are new, think of it as a sorter. You hand it a situation and a few questions, and it gives back an answer for each. Start by downloading the files and trying one sample message.

Under the hood, every question is read out at a special marker token that follows its options, so a single forward pass covers all of them. Extra questions cost little, but they can influence each other, which the card says hurts when the questions are rival candidates. The checkpoint merges several training runs by weight averaging.

State and questions go in as one request
One pass through the 27B model
A marker token follows each question's options
Each marker gives that question's answer

Good to know:

• Hardware: about 52.2 GB of BF16 weights, so plan for a large CUDA GPU
• Long requests are split into parts when they pass the length limit
• License: Apache 2.0
• Cost to run: no paid service named, but you need a capable GPU and a Hugging Face token with access
• Data it reaches: the text and images you send, plus the files downloaded from Hugging Face
• Not for: machines without a CUDA GPU, since the card asks for one

Title: Overall Decision Index 0.3
v1.1 27b: 62.30 points
v1.2 joint requests: 68.09 points
v1.2 one question per request: 67.75 points

On Hugging Face, 13 November 2026

• Task: text classification
• Library: pytorch
• Parameters: 26.1B
• Weights: about 52.2 GB in BF16; about 13 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 5,700
• Likes: 74
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Fine-tunes: 15
• Tags: pytorch, safetensors, qwen3_5, classification, multimodal, custom-code, decider, decision-model, text-classification

Links

• Model: https://huggingface.co/perplexity-ai/pplx-decider-v1.2
• Files: https://huggingface.co/perplexity-ai/pplx-decider-v1.2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download perplexity-ai/pplx-decider-v1.2

Then follow the usage in its model card. Nothing in this kit runs by itself.