AI news · October 13, 2026
Perplexity's pplx-decider v1.1 lifts Decision Index to 61.56
Perplexity's pplx-decider v1.1 raised its Decision Index score from 56.4 to 61.56; the 26B model needs authenticated Hugging Face access and about 49 GiB for weights plus working memory.
- reported Decision Index score
- 61.56
- model parameters
- 26B
- reported weights plus working memory
- 49 GiB
Perplexity's pplx-decider-v1.1-27b is an update to a model built for structured choices rather than open-ended chat. The model card reports a Decision Index score of 61.56, up from 56.4 for the earlier version and above the listed 57.9 score for Jev. The update uses the same Qwen3.8-27B backbone, a separate decision readout and a noncausal attention mode for the evaluated setup. It can return probabilities for choices such as routing a support request to billing, support or sales.
The practical limits are clear: the repository is private, needs authenticated Hugging Face access, and calls for about 49 GiB of weights plus working memory on a CUDA GPU. It is not deployed by an inference provider. That makes this a research tool for teams with the hardware, not a drop-in API. If access is available, compare calibrated probabilities and abstentions on a labelled choice set; do not replace a chat model until the model improves the decision you actually make.
What you can do with it
Open the model card and request access only if you have a suitable GPU and a labelled decision set. Run ten to fifty choices with fixed categories, compare probability calibration and abstention quality with your current model, and record the exact attention and temperature settings. Keep the hosted path unchanged until the local result earns a place.
Our take
A dedicated decision model can be cleaner than forcing a chat model to choose among options, and the score increase is large enough to merit a test. The model is still costly to run and the benchmark is not a business evaluation. Its near-term value is for teams with repeated, labelled routing work and a need for probabilities.
Links Try it
Source: Perplexity / Hugging Face — Made With Models writes the brief; the reporting is theirs.