# Perplexity releases pplx-decider, an Apache-2.0 27B decision model

> pplx-decider-v1 and v1.1 are open-weight 27B models that return probabilities for fixed answers. Built on Qwen3.8-27B.

Published 2026-10-01

Canonical: https://shipped.anth.us/decisions/perplexity-pplx-decider/

Perplexity published pplx-decider-v1-27b on Hugging Face on 1 October 2026 and pplx-decider-v1.1-27b on 5 October, by repository creation date. Both are open-weight decision models under Apache-2.0. Given a state and questions, they return probabilities for the answers rather than text.

The v1.1 card describes a Qwen3.8-27B-based model with a decision head, about 26B parameters and roughly 49 GiB of weights. It needs Python 3.12 or later and a CUDA GPU, and it uses bundled code. Full-attention layers must run in non-causal mode to match the evaluated results. Text plus images must fit in 8,192 tokens, and oversized inputs are rejected.

The card reports a Decision Index of 61.56, against 56.4 for v1 and 57.9 for Jev. It also gives accuracy 0.75 and Brier 0.16. These are from the card itself.

The card text says the repository is private and needs authenticated access. The Hub API lists it as public and not gated. Press coverage mentions a later v1.2 and a smaller Flash variant; that is not covered here.

An open model that beats the hosted incumbent on a public index is worth noting. The index is the maker's choice of yardstick.

## Sources
- https://huggingface.co/perplexity-ai/pplx-decider-v1.1-27b
- https://huggingface.co/spaces/multimodalart/jev-decision-index

## Editor's first pass (AI-assisted)
- Real 3.5: Weights are listed on Hugging Face under Apache-2.0, though the card text mentions authenticated access.
- Useful 3.5: Needs a CUDA GPU with about 49 GiB of weights and custom code, which limits who can run it.
- Notable 3.5: A well-known lab shipping open decision weights, ahead of Jev on its own index.
- Novel 3: Same format as other decision models; the base is Qwen.