---
type: "post"
id: "perplexity-released-decider-v1-3-27b-and-flash-28f495"
slug: "perplexity-released-decider-v1-3-27b-and-flash-28f495"
title: "Perplexity released Decider v1.3 (27B) and Flash v1.3 (4B), open-weight decision models under Apache 2.0."
description: "Perplexity’s Decider v1.3 is a multimodal classification model built on Qwen3.8-27B, with a new Locate mode that returns verbatim text spans, character offsets and scores."
retrieval_nugget: "Perplexity’s Decider v1.3 is a multimodal classification model built on Qwen3.8-27B, with a new Locate mode that returns verbatim text spans, character offsets and scores. Its typed request interface can answer multiple decision questions from one state, making it a local option for teams building structured routing or extraction systems."
published_at: "2026-11-15"
updated_at: "2026-11-15"
record_date: "2026-11-15"
date_kind: "published_at"
topics: ["models"]
entities: ["Perplexity","pplx-decider-v1.3","Qwen3.8-27B"]
related_patterns: ["model-answers-become-task-interfaces"]
novelty_label: "new_fact"
what_changed: "Perplexity has released Decider v1.3 (27B) and Flash v1.3 (4B), open-weight decision models under Apache 2.0."
prior_refs: ["https://newruntime.com/posts/perplexity-s-pplx-decider-v1-2-is-a-c3587d/"]
source_url: "https://huggingface.co/perplexity-ai/pplx-decider-v1.3"
source_urls: ["https://huggingface.co/perplexity-ai/pplx-decider-v1.3"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:perplexity-released-decider-v1-3-27b-and-flash-28f495"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/perplexity-released-decider-v1-3-27b-and-flash-28f495/","markdown":"https://newruntime.com/posts/perplexity-released-decider-v1-3-27b-and-flash-28f495.md","json":"https://newruntime.com/posts/perplexity-released-decider-v1-3-27b-and-flash-28f495.json"}
markdown_url: "https://newruntime.com/posts/perplexity-released-decider-v1-3-27b-and-flash-28f495.md"
---

# Perplexity released Decider v1.3 (27B) and Flash v1.3 (4B), open-weight decision models under Apache 2.0.

## Retrieval answer

Perplexity’s Decider v1.3 is a multimodal classification model built on Qwen3.8-27B, with a new Locate mode that returns verbatim text spans, character offsets and scores. Its typed request interface can answer multiple decision questions from one state, making it a local option for teams building structured routing or extraction systems.

## Summary
Perplexity’s Decider v1.3 is a multimodal classification model built on Qwen3.8-27B, with a new Locate mode that returns verbatim text spans, character offsets and scores. Its typed request interface can answer multiple decision questions from one state, making it a local option for teams building structured routing or extraction systems.

## What happened
Perplexity’s Hugging Face model card describes pplx-decider-v1.3 as the next version of pplx-decider-v1.2. It is a multimodal classification model built on Qwen3.8-27B, with a separate decision readout rather than a full-vocabulary language-model head. The new capability is Locate: a request can ask the model to find text spans matching a label or instruction, and the response includes each span’s text, start and end character offsets, and score. The offsets are exclusive at the end, so the returned text corresponds to the indicated slice of the input state.

The model’s `decide` interface accepts a state and one or more typed questions. The state can be a string, dictionary or list; non-string values are serialized as JSON. Question types include `choice` among 1–255 options, a true-or-false `noul` decision, a score over 2–255 ordered levels, and Locate. A single request can combine multiple decision questions against the same state, and requests can include images for decision tasks. The card’s example classifies a support request by team and urgency. It also says that the model returns probabilities and confidence for relevant decision types, while `logits` exposes raw readout values.

By default, the `separate` mode evaluates each question in its own sequence, so adding or reordering other questions does not change an answer. The `isolated` mode reads the state once and lets each question attend to the state and itself. Perplexity says isolated is cheaper when many questions share a long state, but answers can shift slightly with the question set: on its samples, 1.5–2.2% of top answers differed from separate. The card recommends separate unless the cost of rereading the state matters. Its optional fast path reported 1.7–1.9× speedups over padded batches in separate mode and 9–18× over the reference isolated path, with top-answer agreement ranges of 99.6–99.7% and 99.4–99.8%, respectively, on Perplexity’s validation sample.

The model card specifies Python 3.12 or later and a CUDA GPU with room for about 49 GiB of BF16 weights plus working memory. The repository is gated: users need a Hugging Face token with access. The card sets the default maximum length to 32,768 tokens and says inputs are not truncated. Locate requires nonempty text, no images, and at most 128 questions per request; mixed Locate and decision requests need separate forward passes for each type. The model is tagged Apache-2.0, though the card’s repository access requirement still applies.

## What the archive adds
New Runtime’s earlier coverage described Decider for Mac running pplx-decider-flash-v1.2 locally ([October 24, 2026](https://newruntime.com/posts/perplexity-releases-decider-for-mac-to-run-pplx-25c390/)) and covered pplx-decider-v1.2 as a new version of Perplexity’s decision model ([November 2, 2026](https://newruntime.com/posts/perplexity-s-pplx-decider-v1-2-is-a-c3587d/)). This is a new release in that story, with Locate as a concrete added capability.

## Trend signal
This release supports the shift toward task-specific model interfaces rather than free-form answers. Decider exposes bounded question types and structured results, including offsets for extraction, so software can consume an answer as a typed output. That is a fit with [model answers becoming task interfaces](https://newruntime.com/patterns/model-answers-become-task-interfaces/), though the model card does not establish how broadly teams will adopt this design.

## Why it matters
For a team building a support router, document processor or multimodal triage system, a model that returns a classification and the exact supporting text can reduce the glue between inference and application logic. Character offsets make it possible to attach a result to the original input rather than asking a second generative step to reproduce a quote. Typed questions also make it straightforward to express a small decision schema around a shared case record.

The operational trade-off is significant: this is a large local checkpoint, not a lightweight drop-in classifier. Teams need compatible CUDA hardware, sufficient memory, and time to validate throughput and accuracy on their own workload. The distinction between separate and isolated modes matters too: reusing a long state can save computation, but may make answers sensitive to which other questions are bundled with it. For production systems, that means testing the exact request shape and calibrating probabilities and thresholds on real traffic.

## Limits and uncertainty
Perplexity’s speed, agreement and answer-shift figures are from its own validation samples; the card does not establish that they transfer to a particular application’s data or hardware. The effects of task-specific calibration, thresholds and multimodal inputs also need evaluation in the intended workflow. Although the model card labels the repository Apache-2.0, the model repository requires an authorized Hugging Face token. The release’s operational cost at production scale is not established by the card’s local hardware requirement alone.

## Reader takeaway
Decider v1.3 adds structured span extraction to a decision-model interface, which may help teams turn long text or image-bearing cases into typed outputs. Treat its speed and accuracy figures as starting points, not deployment guarantees: test your request format, hardware, calibration and output quality on your own workload before relying on it.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)