Perplexity has announced the release of its “Decider” model series, a new suite of open-weight classification and decision models designed to provide probability distributions over specific answer options rather than generating text token by token. The release includes two primary models under the Apache 2.0 license: pplx-decider-v1.3 (26B) and pplx-decider-flash-v1.3 (5B).

Model Specifications and Architecture

The larger model, pplx-decider-v1.3, is fine-tuned from the Qwen3.8-27B backbone and requires approximately 49 GiB for its BF16 weights. The smaller, more efficient version, pplx-decider-flash-v1.3, is based on Qwen3.5-4B. According to technical details shared on Reddit, the Flash model was distilled from the 27B version using a KL objective on the teacher’s distributions, requiring roughly 8.5 GiB for its BF16 weights.

Unlike standard generative large language models (LLMs), these decision models utilize bidirectional full-attention layers. Consequently, developers must use the specific pplx_decider package to ensure accurate outputs, as standard causal inference will not reproduce the intended results. Both models support text and image inputs and are accessible via Perplexity’s API and OpenRouter.

New “Locate” Functionality

The v1.3 update introduces a new question type called locate. This feature allows the model to extract verbatim spans from a given state, providing character offsets for tasks such as Named Entity Recognition (NER), Personally Identifiable Information (PII) detection, and extractive question answering.

Perplexity notes that the locate functionality operates via a separate LoRA adapter and a span head that remains unmerged with the main model to prevent performance degradation in classification tasks. While the standard decision models support both text and images, the locate feature is currently limited to text-only inputs and can handle up to 128 locate questions per request.

Performance Benchmarks

Perplexity has reported several performance metrics comparing the Decider series to existing industry tools:

  • Extraction Tasks: In 17 out-of-domain tasks (including NER and PII), the v1.3 model achieved an average score of 69.0, while the Flash version scored 63.6. These figures compare to 43.3 for Liquid d1 and 33.6 for GLiNER large v2.5.
  • Multilingual Classification: Across eight languages, the v1.3 model scored 88.16% and the Flash model 86.50%, compared to 85.41% for TypeSafe’s Jev 1.14.
  • Reranking: The models achieved an 83.99 nDCG@10 across 72 retrieval tasks, a 1.81 increase over Qwen3-Reranker-8B.

According to the Artificial Analysis Decision Index, the v1.3 model sits at a cost of approximately $68 per 1M tasks with a score of roughly 70, while the Flash model costs approximately $34 per 1M tasks with a score of 64.

Implementation Considerations

Perplexity has advised users that returned probabilities should be calibrated and thresholded based on specific datasets. Additionally, because the locate adapters are unmerged, requests that combine both classification and extraction tasks will require extra forward passes, increasing computational requirements.