By No Code MBA · Facts checked
- Provider
- Perplexity
- Announcement date
- Status at last check
- Generally available
- Model family
- pplx-decider
- Preceding model
- pplx-decider-v1.1-27b
- Capabilities
- Text, Image
What changed
Version 1.2 answers several questions about one input in a single pass, where v1.1 answered one question per forward pass, returning choices, yes/no probabilities or scores rather than free text. The main model is 27B on a Qwen3.8-27B backbone, and the Flash model is about 4B to 5B, distilled from it on Qwen3.5-4B. Perplexity reports a Decision Index overall score of 68.09 for v1.2 versus 62.30 for v1.1, and 58.88 for Flash, plus a Vision Decision Index of 77.43 for v1.2.
Access and availability
Both models are on the Perplexity Decisions API at POST https://api.perplexity.ai/v1/decisions as pplx-decider-v1.2 and pplx-decider-flash-v1.2, with a 256K token context. Apache 2.0 weights are listed on Hugging Face under perplexity-ai, though the model cards say the repositories are private and need an authorized token.
Official access documentationPricing
pplx-decider-v1.2: $0.02 per million input tokens; pplx-decider-flash-v1.2: $0.01 per million input tokens; output is free
Official pricingWhat you can build
Try both sizes on a small set of your own routing or classification cases, then review probability calibration and uncertain decisions with a human before automating actions.
Limitations
Training used sequences up to about 32K tokens, so longer inputs are only spot-checked, and v1.1 prompts are not compatible. Perplexity says grouping questions can hurt accuracy when candidates compete independently, and Flash is recommended for one question per request. Self-hosting the 27B model needs about 52 GiB of GPU memory for BF16 weights.