Perplexity released Decider v1.3 and Decider Flash v1.3 on November 12, 2026, adding exact text-span extraction to models built for classification, routing, moderation, reranking and evaluation. Flash is a smaller model distilled from the larger Decider.
Both models support text and vision tasks. Their weights are available on Hugging Face under Apache licensing, with hosted access through OpenRouter and the Perplexity Decisions API, according to the company’s announcement.
The practical distinction is their output. A chatbot generates an answer token by token. Decider returns probabilities over supplied choices, or extracts matching text through its new locate feature. That makes it worth evaluating wherever an application needs a structured decision without an accompanying explanation.
Perplexity reports strong benchmark results, including 91.6% mean accuracy across 63 classification tasks. Those figures come from the provider’s evaluations, so the useful question for developers is how closely those tasks resemble their own workloads.
Decision Models Turn Questions Into Probabilities
Perplexity describes the category plainly: “A decision model is a general-purpose classifier.”
A request contains a shared input, called the state, plus questions about that input. Each classification question supplies its possible answers. Decider returns a probability distribution over those answers, supporting three formats:
| Question type | Example application | What the application receives |
|---|---|---|
| Yes/no | Does this tool response contain a prompt-injection attempt? | Probabilities for yes and no |
| Multiple choice | Should this request go to billing, support or sales? | A probability for each supplied category |
| Ordered score | How relevant is this document to a search query? | Probabilities across ordered relevance levels |
These are illustrative applications, consistent with the tasks described in Perplexity’s documentation.
Several questions can share one input. A document-and-summary pair, for example, could be checked for faithfulness and intended audience in the same request. Ordinary decision questions share computation in a single forward pass, avoiding token-by-token answer generation.
That design sits between a narrow, task-specific classifier and a generative LLM. Developers can describe new tasks in natural language while keeping the response constrained to choices their software can consume.
I’d see the strongest initial use case in systems that already discard most of an LLM’s response after extracting a label. A structured probability distribution gives those systems something more useful: a basis for thresholds, escalation and ranking. The probabilities still need validation against real outcomes before a team treats them as reliable confidence estimates.
How Version 1.3 Changes the Model
Decider starts with a generative LLM backbone, then adapts its output machinery for decisions.
Each question includes labeled answer options and a special <|marker|> token. A learned linear output layer reads the model’s representation at that position and produces scores for the options. Those scores become probabilities. Perplexity initializes this decision head from the base model’s output layer, retaining the rows associated with predefined option tokens.
Training converts hundreds of public datasets into this shared state, question, options and target format. Tasks include sentiment analysis, intent detection, moderation, entity matching and ratings.
Version 1.3 also makes the backbone’s full-attention layers bidirectional, letting tokens incorporate context from the entire input. Perplexity says its ablation studies found this improved performance.
The interaction between questions changes, too. In v1.2, the state and all questions could attend to one another, so adding or changing one question could affect another answer. Version 1.3 isolates questions from direct attention to each other while allowing the shared state to attend to them, according to Unite.AI’s technical account. The design aims to improve answer stability; it is not a guarantee that changing a request leaves every prediction unchanged.
Flash learns from the larger model’s probability distributions through distillation. Its training minimizes the difference between student and teacher outputs using KL divergence, a measure of how two distributions differ. That preserves more information than training solely on a winning label, because the teacher’s relative uncertainty across choices also contributes to learning.
Locate Adds Extraction, With a Serving Tradeoff
Locate extracts text that already exists in the input. A question can describe a field with a short label, such as location, or give an instruction to extract the city where a job is based.
The response contains matching text plus its start and end character positions. No detected match produces an empty list. Perplexity identifies named-entity recognition, personal-information identification and extractive question answering as intended uses in its release post.
For a job listing, an application could ask whether senior experience is required and extract the listed skills. For privacy screening, it could classify a document and locate potentially sensitive passages.
Character positions matter because they connect a prediction to a specific part of the source. An application can highlight a match, send it for review or use it in a redaction workflow without searching for a generated paraphrase.
One Request Does Not Always Mean One Forward Pass
Locate has three important serving constraints, reported by Unite.AI:
- It is text-only, despite the models’ broader image support.
- A request can contain up to 128 locate questions.
- Mixing locate and decision questions requires separate forward passes, adding latency compared with requests containing only one type.
The extraction adapter remains separate because merging it into the model caused regressions on ordinary decisions. It activates for locate requests.
That detail changes how developers should interpret the shared-input design. Combining extraction and classification simplifies the application’s request structure, but it does not fully combine their computation. For an interactive pipeline, I’d measure mixed-request latency separately from classification-only latency before choosing the architecture.
Extraction also needs its own quality checks. An empty list means the model found nothing, not that sensitive information is certainly absent. And offsets must be applied to the same source text: preprocessing that changes character positions can invalidate a later highlight or redaction.
Hosting Options and the Flash Tradeoff
Perplexity offers three distribution routes: downloadable Hugging Face weights, OpenRouter and its own Decisions API. Local deployment gives developers control over execution and infrastructure; hosted access avoids operating the model themselves.
The following deployment details are reported by MarkTechPost, with approximate weight footprints also described by Unite.AI:
| Detail | Decider v1.3 | Decider Flash v1.3 |
|---|---|---|
| Model role | Larger model | Smaller distilled model |
| Perplexity API model ID | pplx-decider-v1.3 | pplx-decider-flash-v1.3 |
| Supported inputs | Text and images; locate is text-only | Text and images; locate is text-only |
| Approximate BF16 weight footprint | 49 GiB | 8.5 GiB |
| Reported Perplexity API input price | $0.02 per million tokens | $0.01 per million tokens |
| Reported API output price | Free | Free |
BF16 is a 16-bit numerical format used to store model weights. Those footprints are not total GPU-memory requirements: inference also needs working memory. The documented local setup requires a CUDA GPU, making Flash substantially easier to accommodate, although its weight size alone cannot determine whether a particular card is sufficient.
The published local loader uses the pplx_decider package and Decider.from_pretrained. Deployment must also preserve the model’s noncausal attention behavior; this is not simply a chat model that can be dropped into any generative serving configuration.
Frequently Asked Questions
4 questions
1Is Decider v1.3 a Chat Model?
No. It returns probabilities over supplied answer choices, ordered scores or extracted text spans. It is designed for structured decisions, not conversational prose.
