Perplexity Drops Open-Weight pplx-decider to Replace JSON Parsing With Typed AI Decisions
Perplexity's new 27B and 4B decision models return calibrated probabilities instead of text, answer multiple questions in one pass, and cost two cents per million tokens.
- Perplexity released pplx-decider-v1.2 (27B) and flash-v1.2 (4B), open-weight multimodal decision models under Apache 2.0.
- Returns calibrated probabilities for typed questions instead of generated text; no JSON parsing required.
- Joint request format answers all questions in one forward pass, up to 5x faster inference.
- Pricing: $0.02 per million input tokens for 27B, half for flash; output tokens are free.
- 27B scores 68.09 on Decision Index (up from 62.30); flash tops the 4B tier at ~60.
- Group questions for multi-facet tasks; send separately for independent candidate ranking.
Perplexity has released two multimodal models built for classification, routing, moderation, and rubric grading. pplx-decider-v1.2 has 27 billion parameters, while pplx-decider-flash-v1.2 has 4 billion. Both accept text, structured data, and images, then return typed answers with calibrated probabilities.
Both checkpoints use the Apache 2.0 license, allowing developers to download, modify, and self-host the weights. Perplexity also serves them through its Decisions API and third-party model gateways.
| Model | Parameters | License | Primary use |
|---|---|---|---|
pplx-decider-v1.2 |
27B | Apache 2.0 | Higher-accuracy decisions |
pplx-decider-flash-v1.2 |
4B | Apache 2.0 | Lower-cost, lower-latency decisions |
Typed outputs remove JSON repair
Many production LLM tasks have a fixed answer space: choose a support team, score a response against a rubric, flag unsafe content, or decide whether to escalate. Decider models expose those tasks directly, allowing application code to threshold probabilities without parsing a generated chat response or repairing malformed JSON.
A request contains a state and one or more questions. The state can include text, structured fields, or images. Each question uses one of three types: choice for named options, noul for yes-or-no judgments, and score for ordered grades. The API returns option probabilities, an expected answer, and a confidence value.
decider.decide({
"state": "My integration keeps failing with a 401 since we rotated our API key.",
"questions": {
"team": {
"type": "choice",
"criteria": {
"billing": "Charges and refunds",
"support": "Technical integration errors",
"sales": "Questions about buying a product"
}
},
"urgent": {
"type": "noul",
"instructions": "Is the customer blocked right now?"
},
"frustration": {
"type": "score",
"criteria": ["calm", "mildly annoyed", "frustrated", "angry"]
}
}
})This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
