What pplx-decider is
pplx-decider-v1-27b is Perplexity’s System One model, released on 1 October 2026 with the Decisions API. Perplexity describes it as a multimodal decision model trained to output a probability distribution over a fixed set of answers instead of text. You send a state and up to 128 named questions, and it answers all of them in one call.
It is listed as commercial because Perplexity’s own generally available API is the main way to use it. The weights are also on Hugging Face under Apache 2.0, ungated, on a Qwen3.8-27B base. The card says 26 billion parameters, though the name says 27b.
What it returns
The request is POST https://api.perplexity.ai/v1/decisions with model set to pplx-decider-v1-27b. A Noul returns a yes probability from 0 to 1. A Choice takes 1 to 255 options and returns a probability for each. A Score takes up to 10 ordered levels and returns a probability-weighted average. The state can be text, JSON, or a list that includes base64 images. The response is JSON, and output tokens are not billed.
What it is good at
Perplexity’s docs include a support-ticket triage example, which maps to support inbox triage. Routing a request to a model or tool fits intent and model routing. The model card reports 88.80 on RAGTruth, a test of whether an answer is supported by its sources, which is closest to citation verification. The context window is 262,144 tokens, much larger than most hosted rivals, and it accepts images.
The card reports 85.71% across 11 benchmarks, against 84.51% for Jev and 74.76% for the base Qwen3.8-27B. Jev scores higher on 6 of the 11. Perplexity measured its own model through its API and does not say how it ran Jev, so these are vendor-run.
The independent Decision Index lists it, checked on 2 October 2026, under the engine id autojev-27b, which the site’s guides earlier called AutoJev-27B. It scores 56.40 against Jev’s 57.91 on the 0 to 100 headline score, with a calibration error of 0.018 against Jev’s 0.074.
What it is not for
It writes no text, so it cannot explain an answer or draft a reply. Perplexity publishes no calibration data of its own. There is no response-time guarantee, and latency grows with input size. The rate limit of 10 requests per second per organization caps high-volume use. Perplexity’s Python and TypeScript SDKs do not document a decisions method, so the quickstart calls the endpoint with httpx or fetch.
Access
The Decisions API is generally available now with a Perplexity API key. Self-hosting needs Python 3.12 or newer and a CUDA GPU with about 49 GiB for the weights plus working memory. The price was $0.04 per million input tokens on 2 October 2026, and Perplexity says it plans to lower it.
Specifications
| Question types | ChoiceScoreNoul |
| Max Choice options | 255 |
| Score levels | Up to 10 |
| Questions per call | 128 |
| Total context | 262,144 tokens |
| State budget | Not documented |
| Rate limit | 10 requests per second per organization, on all plans. Going over returns HTTP 429 with a Retry-After header. |
| Endpoint | POST https://api.perplexity.ai/v1/decisions |
| SDKs |
A request must stay under 262,144 input tokens and the body under 32 MiB, which covers the state and all questions together. A Choice takes 1 to 255 options and a Score up to 10 levels. State can be a string, an object or an array. Images go in the state array as base64 PNG, JPEG or WebP data URLs, up to 2,048 tiles of 32 by 32 pixels, for example 1440 by 1440.
Versions
- pplx-decider-v1-27b, 1 Oct 2026, The only model on the Decisions API. Open weights on Hugging Face under Apache 2.0, ungated, on a Qwen3.8-27B base, in BF16 Safetensors. The card says 26 billion parameters while the model id says 27b. Srinivas's launch post calls it pplx-decider-27b. Release notes
Use cases
What people use pplx-decider for, one page per pattern.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Intent and model routing with System One models
One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.
Citation verification with System One models
An answer with citations is only as good as the citations. Match each quote against the source in code to catch fabrications, then ask Jev one Choice question about how the surrounding section relates to the claim. Confidence decides which verdicts a human reviews.