pplx-decider-v1.1-27b

pplx-decider-v1.1-27b

Picks the best option for a decision from text and images, such as which team should handle a request, with a probability for each choice, for developers who build routing and review workflows.

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-v1.1-27b is a 27B decision model from Perplexity, built on Qwen3.8-27B. You give it a situation as text, optionally with images, plus a question and a set of choices, and it returns a probability for each choice. The v1.1 release lifts v1's overall rating on the Decision Index, with 56.4 before and 61.56 now.

Who it suits: It suits developers and operations teams who want a model to route requests, grade items or make multiple-choice calls inside a workflow.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Customer support | Support leads | Send each request to the right team | Choose between billing, support and sales for a message
Operations | Workflow builders | Make structured calls about incoming items | Pick one label for a ticket from a list of criteria
AI research | Evaluation researchers | Study decision models on the Decision Index | Compare it with the v1 model

If you need a chat model that writes text, you can skip it, since this one returns decision probabilities.

Where to use it:

• Python scripts with PyTorch on a CUDA GPU
• Workflows that route or grade items from text and images
• Windows or Linux with Python 3.12 or newer

Title: Jobs it does
Use case | Who | What they get back
Routing | Support and ops teams | A probability for each team or label
Multiple-choice calls | Developers | A distribution over the options you define
Image-aware decisions | Workflow builders | Choices informed by attached pictures

How to set it up: Plan for Python 3.12 or later, an Nvidia CUDA card that can hold about 49 GiB of weights with space to spare, and a Hugging Face login. The card calls the repository private, so log in with an account that has access.

• Python 3.12 or newer
• A CUDA GPU with plenty of memory
• A Hugging Face login with access to the repository
• The pinned packages from the repository's requirements.txt

1. Fetch the repository, then install the exact package versions listed in its requirements file.
2. Run the sample script. It loads the model and the inference code that ships with it.
3. Replace the example question and choices with your own.

pip install -U huggingface_hub
hf download perplexity-ai/pplx-decider-v1.1-27b

How to use it: Build a question with a type, instructions and a set of named criteria. Pass it with a state, which is the situation text, then call predict to get probabilities. The answer helper picks the result from them. Use the included inference code, because default causal attention does not reproduce the evaluated setup.

import sys
from pathlib import Path
from huggingface_hub import snapshot_download

checkpoint = Path(snapshot_download("perplexity-ai/pplx-decider-v1.1-27b"))
sys.path.insert(0, str(checkpoint / "source" / "src"))
from autojev.model import DecisionModel, answer

model = DecisionModel(checkpoint, device="cuda")
question = {
    "type": "choice",
    "instructions": "Which team should handle this request?",
    "criteria": {
        "billing": "Charges and refunds",
        "support": "Technical integration errors",
        "sales": "Questions about buying a product",
    },
}
row = {"state": "My integration keeps failing. Please help.", "question": question}
probabilities = model.predict([row])[0]
print(answer(question, probabilities))

For images, add them to the row as file paths, PIL images or base64 data URLs. The processor keeps the aspect ratio and caps the size with max_image_pixels, which you can change.

model = DecisionModel(checkpoint, device="cuda", max_image_pixels=1600 * 1000)

Example: A message says an integration keeps failing and asks for help. The question asks which team should handle it, with billing, support and sales as choices. The model returns a probability for each team, and the helper reports the best fit.

For newcomers and experts: Think of it as a classifier that reads a situation and your options and tells you how likely each option is. A first small test is the sample script with your own message and choices.

The checkpoint is a Qwen3_5Model backbone with a separate decision head, and it has no full-vocabulary language-model head. Full-attention layers run in noncausal mode, while the linear-attention layers keep their normal behavior. If you work with raw logits yourself, scale them by the stored temperature one time only, then normalize across the options that apply to that decision.

Situation text and images
Noncausal backbone reads them
Decision head scores the options
Probabilities per choice

Good to know: Inputs must fit in 8,192 tokens, and anything longer is rejected, not truncated. The model saw images no larger than 512 by 512 pixels in training and testing, so bigger pictures are outside what it learned. Most of the gain over v1 came from removing the causal mask and adding more training data.

Title: Overall Decision Index
Jev: 57.9 points
pplx-decider-v1: 56.4 points
pplx-decider-v1.1: 61.56 points

• License: Apache 2.0.
• Cost to run: none named, apart from a CUDA GPU.
• Data it reaches: the files come from Hugging Face behind a login, and the model then runs on your own GPU.
• Not for: chat or text writing, or servers that need a standard language-model head without a custom export.

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: 4,065
• Likes: 182
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 4
• Fine-tunes: 3
• Tags: pytorch, safetensors, qwen3_5, classification, multimodal, custom-code, decider, decision-model, text-classification

Links

• Model: https://huggingface.co/perplexity-ai/pplx-decider-v1.1-27b
• Files: https://huggingface.co/perplexity-ai/pplx-decider-v1.1-27b/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-v1.1-27b

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