Perplexity: Decider V1.1 27B

Decider V1.1 27B is Perplexity’s updated decision checkpoint. It evaluates supplied content against a defined question and produces probabilities over the valid answers. Its native artifact has a decision readout rather than a full-vocabulary language-generation head. The useful result is a bounded decision distribution and its interpreted answer, not an open-ended explanation or a general chat response.

What changes in the V1.1 edition

Perplexity identifies this model as an update of Decider V1 27B that retains the same backbone, with changes to the attention setup and more training data. Its backbone is Qwen3.8-27B. The V1.1 edition should retain its own name and configuration rather than being described as an interchangeable copy of the earlier checkpoint.

State, question and valid answers

The published example supplies a row containing a state and a typed choice question. The question includes instructions and criteria whose keys are the semantic answers and whose descriptions define when each applies. The model predicts probabilities, then the producer’s answer helper interprets those probabilities for the question. The inference example can process a list of rows, with one decision tied to each row’s evidence and question.

For request routing, provide the actual message and relevant context, then define the departments or categories available. An illustrative task might distinguish an integration failure, an access problem and a request for product information. Make each description precise enough to settle a borderline case, and retain the option definitions with the result. A category distribution can then support a routing decision or a review step when alternatives remain close. The model’s chosen result does not itself execute the workflow action.

Text and image evidence

The native implementation accepts images alongside the state and question, using image-file inputs, in-memory images or encoded image data. For a document image or screenshot, supply the visual evidence and ask a closed question whose options describe the relevant outcomes. Identify the visible feature that matters and provide the text context needed to interpret it. For example, classify whether a pictured item belongs to one of the supplied product categories rather than asking the decision model to write a complete product description.

The processor preserves aspect ratio during resizing and rounds image sides to multiples of 32 pixels. The documented processed area begins at 65,536 pixels and is bounded by a configurable maximum; its default is 4,194,304 pixels, or 2048 by 2048. Each 32-by-32 patch contributes one visual token, so that default permits up to 4,096 visual tokens for an image. These figures describe the producer’s native image preparation, not a general application upload contract.

Combined input limits and resolution scope

The documented prepare step requires the text and all images in one decision to fit within 8,192 tokens. Oversized input is rejected rather than silently truncated. Several images share that budget with the written state and question. Reduce irrelevant source text or the image-size cap when preparing a combined decision, while retaining the detail that determines the answer.

The training and reported evaluation setup capped images at 512 by 512, corresponding to 262,144 pixels and up to 256 visual tokens. The configurable inference cap can retain more detail, but the card warns that higher resolution is outside that training distribution and uses more memory and processing time. A larger image setting is therefore not a guarantee of a better decision. Keep the input preparation consistent with the question and inspect whether the needed visual feature is represented clearly.

Attention, readout and probability calibration

The native checkpoint’s full-attention layers use a saved noncausal attention mode, while its linear-attention layers retain their own behavior. The maker states that default causal inference does not reproduce the evaluated setup. Its prediction implementation applies the saved calibration temperature; if logits are used directly, that temperature is applied once and the probabilities are normalized over the current valid candidates. These details belong to the exact decision profile and should be preserved when interpreting its numerical outputs.

The separate decision readout also distinguishes this artifact from a generative export. A conversion to a vocabulary-based generation interface needs an explicit mapping and must preserve the attention behavior. Prepare closed questions for the native decision profile; use a suitable generation model when the task needs chat, code, explanations or tool requests.

A useful final artifact keeps the state, question, allowed answers and resulting distribution associated. Define the question around evidence actually present, review ambiguous cases against that evidence, and use labeled examples from the intended task to decide how numerical outputs should affect routing or classification. The model supplies decision probabilities; the criteria and the surrounding application determine how those probabilities become a useful action or recorded label.