Perplexity releases a 27B model for fixed-choice decisions

Perplexity has released pplx-decider-v1-27b, a model that maps text, structured data, or images to probabilities over developer-defined answers. The company published the weights under Apache 2.0 and launched a hosted Decisions API priced at $0.04 per million input tokens, with output unbilled.

The model targets workflows that ask a chat model for a label and then parse its response. Classification, routing, moderation, and rubric-based grading can use numeric probabilities, explicit answer sets, and application-defined thresholds.

A fixed answer space

The API receives content in a state field and evaluates it against one or more named questions. The state can contain text, JSON, or images encoded as base64 data URLs. The API supports three question types:

• noul: Evaluates a yes-or-no question and returns the probability of yes from 0 to 1. For an exhaustive binary decision, the probability of no is the complement.
• choice: Selects among developer-defined options and returns each option’s probability along with the leading choice.
• score: Evaluates an ordered rubric and returns a probability-weighted expected level.

A single request can apply several questions to the same state. A support request, for example, could be classified by destination team, checked for escalation, and scored for severity without resending the underlying text in separate calls.

The repository’s inference helper exposes the same pattern for local use:

from inference import Decider

model = Decider.from_pretrained("perplexity-ai/pplx-decider-v1-27b")
result = model.predict(
    "My Stripe integration keeps failing. Please help ASAP.",
    {
        "type": "choice",
        "instructions": "Which team should handle this request?",
        "criteria": {
            "billing": "Charges and refunds",
            "technical_support": "Integration errors",
            "sales": "Questions about buying a product",
        },
    },
)
print(result)