Most "classify this" pipelines ask a chat model for a label and then parse whatever comes back. This model skips the prose entirely. You send a state (text, JSON or images) and a set of typed questions; you get calibrated probabilities back. Version 1.3, released 9 November 2026, adds a fourth question type that turns it into an extractor as well.
New in this release: locate
Earlier decider versions could pick an option, answer yes or no, or place something on a scale. locate returns verbatim spans from the state, each with start and end character offsets and its own score, so state[start:end] reproduces the text exactly. Typical uses: pull the clause that triggers a refund, the sentence that contains a policy violation, or every product name in a review.
Constraints worth noting: locate reads text only (no images), needs a non-empty text state, and allows up to 128 questions per request. Only the v1.3 generation serves it.
The four question types
| Type |
Returns |
choice |
One of 1–255 named options, per-option probabilities, confidence |
noul |
Probability that a statement is true, with optional criteria |
score |
Expected level on a 2–255 step scale, probabilities, legend |
locate |
Matching spans with offsets and scores |
Several questions about the same state can go in one request, and each is answered in isolation from the others.
Pricing and access
Through Perplexity's Decisions API (POST /v1/decisions) the model costs an indicative $0.02 per million input tokens, and output is free. pplx-decider-latest now resolves to it. The API's optional web grounding mode runs a fast search per question (up to 100) for $0.001 each, and fetching a URL passed as state costs another $0.001; the Python and TypeScript SDKs do not expose the mode field yet, so set it in the raw request body. Confirm current rates on your own account.
The weights are tagged Apache-2.0 and built on Qwen3.8-27B, but the repository was private and gated when we checked. Expect to request access. Self-hosting in BF16 needs a CUDA GPU with roughly 49 GiB plus working memory and Python 3.12.
Migrating from v1.2
Two warnings, both practical.
First, the prompt protocol changed. v1.3 uses its own format (protocol version 4), so prompts produced by v1.1 or v1.2 builders must not be reused; build requests with the bundled library. Second, probabilities shift between versions. Any threshold you tuned on v1.2, such as "flag if noul > 0.82", needs revalidating against a labelled sample before you flip the alias in production. The explicit v1.2 ID remains available if you need time.
Size and length: conflicting figures
The name says 27B and the base is a 27B model, but the weight metadata counts 26B parameters. The self-hosted default max_length is 32,768 tokens, much shorter than the 262,144 default the previous generation shipped with; inputs over the limit raise an error rather than being truncated. If you feed long documents, test your longest real input on the hosted API before assuming the old headroom still applies.
Should you use it?
If your workload is a fixed set of decisions over many items (moderation, routing, triage), compare it against your current LLM-as-judge on 500 labelled examples. Look at calibration, not just accuracy: plot predicted probability against observed rate. A cheaper, faster sibling, pplx-decider-flash-v1.3, shares the schema. For other approaches see structured decision models and zero-shot classification; Perplexity's search models are on the provider page.