pplx-decider-flash-v1.2

pplx-decider-flash-v1.2

A small Perplexity decision model built on Qwen3.5-4B that reads text and images and answers choice, yes-or-no and score questions with probabilities, for fast routing and triage.

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-flash-v1.2 is a small decision model from Perplexity. You give it a piece of text, sometimes with images, plus one or more questions. It answers each question with a choice, a yes-or-no probability or a score on a scale you define.

It is the compact sibling of pplx-decider-v1.2, at about one seventh of the size, and it is built on Qwen3.5-4B. A larger 27B model acted as the teacher, and the small model learned the teacher's whole spread of probabilities, not only its final picks. Several training runs were merged by averaging their weights.

Who it suits: It suits developers who sort or rate requests inside an app and care about speed or cost more than the last points of accuracy. The record lists 4.5B parameters and about 9.1 GB in BF16.

Industry | Who uses it | How they use it | Example
Customer support | Support teams | Route incoming tickets to the right team | Sending a payment problem to billing
Operations | Platform engineers | Flag requests that need action right now | Marking a blocked customer as urgent
Community | Moderators | Rate the tone of a message on a scale | Scoring a note from calm to angry
Software | Developers | Add a cheap decision step to an AI pipeline | Picking a tool before an agent acts

Skip it if you need the best accuracy, since the 27B sibling scores higher, or if you have no CUDA GPU.

Where to use it:

• Python 3.12 or newer, on a machine with a CUDA GPU
• Routing, triage and scoring steps inside apps and agents
• Windows, macOS or Linux, with Python, for the file download step

Use case | Who | What they get back
Request routing | Support teams | The chosen option, a probability for each option and a confidence
Yes-or-no checks | Developers | A probability that the answer is true
Scoring | Analysts | An expected level on your scale, with probabilities
Picture-aware decisions | Developers | Choices that also look at images up to 1.6 MP

How to set it up: You need a few things in place first.

• Python 3.12 or newer
• A CUDA GPU with about 9 GiB free for the BF16 weights, plus extra space to run it
• A Hugging Face token with access to the repository, which the card says is private
• The requirements file from the repository, which pins torch, transformers and related packages

1. Install the Hugging Face hub package with pip and authenticate with your token.
2. Download the repository with the hf command and install its requirements file.
3. Load the checkpoint with the Decider class from the repository's source folder.

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

How to use it: You call the decide function with a request. A request holds the text to judge, one or more questions and optional images. Every question carries a type, and the model sees your options in the order you list them.

Tool | What it does
Decider | Loads the checkpoint onto a CUDA device
decide | Takes one request, or a list of them, and returns answers in the same order
requirements file | Pins the packages the model needs

Question type | What it returns
choice | The chosen option, a probability per option and a confidence
noul | The probability that the answer is true
score | An expected level, probabilities and a confidence

Example: A customer writes that checkout fails after entering a card. You ask which team should take it, billing or support. You also ask whether the customer is blocked. The reply names the team and gives a yes probability you can compare with your own threshold.

For newcomers and experts: Newcomers can think of it as a sorting step that also tells you how sure it is. Start with one question and one short text, then read the probabilities before you act on them.

When a task has several questions, sending them one at a time scored higher for this model on most tests. The 27B model behaves the other way round. Distillation also helped on decision tasks held out from training. The vision score is 72.53 against 77.43 for the larger model.

Title: Overall Decision Index, from the card
pplx-decider-v1.2: 68.09 points
Flash one question per request: 60.05 points
Flash joint requests: 58.88 points

Write the text and your questions
Call the decide function
Read the answers and probabilities
Route or act on the result

Good to know:

• License: Apache 2.0
• Cost to run: none named, but you need a CUDA GPU
• Data it reaches: it downloads files from Hugging Face with your token and reads the text and images you pass in
• Not for: top accuracy when you can run the 27B sibling, writing text, or anyone without access to the private repository

On Hugging Face, 13 November 2026

• Task: text classification
• Library: pytorch
• Parameters: 4.5B
• Weights: about 9.1 GB in BF16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 7,052
• Likes: 78
• License: apache-2.0
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 13
• Fine-tunes: 1
• Tags: pytorch, safetensors, qwen3_5, classification, multimodal, custom-code, decider, decision-model, text-classification

Links

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

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