pplx-decider-v1-27b
Perplexity's open decision model: Qwen3.8-27B fine-tuned with a 255-option decision readout in place of the text head. Reads text, JSON and images and returns probabilities for yes/no, choice and rubric-score questions. Also served as Perplexity's hosted Decisions API.
The readout has a temperature fitted on a separate calibration split, but no calibration error is published. The hosted API (POST /v1/decisions) takes 1 to 128 questions per request, up to 255 options per choice and 1 to 10 score levels, under 262,144 input tokens, in Perplexity's own request format (not a Jev endpoint). There is no lm_head, so the repo's inference code is needed. Perplexity reports 85.71% over its 11-benchmark panel against 84.51% for Jev, measured through its API. v1.1 (2026-10-05, perplexity-ai/pplx-decider-v1.1-27b) runs the full-attention layers noncausally and adds tasksource data; its card cites the third-party Decision Index board at 61.56 against 56.4 for v1. Weights Apache-2.0 with a NOTICE (derived from Qwen3.8-27B); the code under source/ is MIT.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 26.1B |
| Base model | qwen/qwen3.8-27b |
| Maker | Perplexity |
| Released | 2026-10-01 |
| License | apache-2.0 |
| Reported accuracy | 85.7% |
| Reported latency | under 2 s for a few hundred input tokens, 5 s at about 90k, 23 s near the 262,144-token limit (Perplexity's tests, 30 Sep 2026) |
Hosted API
Served by Perplexity — $0.04/MTok input, $0/MTok output. Get access · API docs.
Get the weights
pip install systemonemodels
systemone pull perplexity/pplx-decider
The files are served from the maker's Hugging Face repository, perplexity-ai/pplx-decider-v1-27b, and verified against the checksums recorded here.