Perplexity open-sources pplx-decider v1.2 decision models under Apache 2.0: a 27B (Decision Index 68.1, up from 62.3) and a 4B Flash, all questions in one pass for up to 5x faster inference, at 2 and 1 cents per million tokens

Why it matters

【Key points】

Perplexity's developer account announced its next open-weight multimodal decision models on Oct 16 at 18:42 UTC: pplx-decider-v1.2 (27B, fine-tuned from Qwen3.8-27B) and pplx-decider-flash-v1.2 (4B, Qwen3.5-4B distilled from the 27B models), both Apache 2.0 on Hugging Face. CEO Aravind Srinivas added pricing: 2 cents per million tokens for the 27B model and 1 cent for Flash.

Decision models generate no text. They read a state (text or images up to 1.6 MP) plus questions and return choices, yes/no probabilities or scores with confidence, the routing, classification, moderation and reranking calls inside agents. The main v1.2 change is that all questions in a request are answered in one sequence; Perplexity says inference is up to 5x faster. Per the model cards, the 27B model's Decision Index 0.3 overall rises from 62.30 (v1.1) to 68.09 and vision from 73.22 to 77.43; Flash scores 60.05 and 72.53. Perplexity says the 27B ranks #2 overall on the Hugging Face Decision Index and Flash is the top 4B model.

【Verifiable】

Model card (27B): https://huggingface.co/perplexity-ai/pplx-decider-v1.2

Model card (4B Flash): https://huggingface.co/perplexity-ai/pplx-decider-flash-v1.2

Official post: https://x.com/perplexitydevs/status/2111166029539758120

CEO post: https://x.com/AravSrinivas/status/2111181956444827695

【Our read】

1. Decision models are becoming their own category: Liquid AI shipped d1-ultra this week, and Perplexity has gone v1, v1.1, v1.2 in about two weeks. The fight is over cents-per-million pricing and calibrated probabilities.

2. Discount the gains: the card says this version adds several benchmarks' official training splits (decontaminated) to training, so part of the jump over v1.1 is in-domain data, not pure capability.

3. For builders: Apache 2.0 and a 4B model that fits one GPU make it practical to move many "which one / yes or no" calls off large models, cutting cost and latency.