Perplexity Open-Sources Decider 1.3, Adding Text-Span Extraction

What happened

A decision model is a general-purpose classifier: given an input and a set of questions, each with possible answers, the model returns a probability for every answer in a single forward pass, supporting yes/no, multiple-choice and score questions. Perplexity says such tasks were often handled by generative models even though they do not require token-by-token text generation, and decision models share computation across questions instead of decoding token by token.

The main addition in v1.3 is locate: given a field name or an instruction, it extracts the exact matching text span from the input and returns its start and end character positions, for named-entity recognition, personal-information identification and extractive question answering; no match returns an empty list. Locate is text-only and capped at 128 locate questions per request.

The two models differ in size: Decider 1.3 is built on Qwen3.8-27B, with roughly 49 GiB of BF16 weights; Decider Flash 1.3 is built on Qwen3.5-4B, about 8.5 GiB, distilled from the larger model with a KL-divergence objective. Both are available through OpenRouter and Perplexity’s Decisions API, with a 262,144-token context; the free tier on OpenRouter runs until Nov 19.

The evaluation numbers

Across 63 classification tasks, Decider 1.3 achieves 91.6% mean accuracy, exceeding Perplexity’s production classifiers by 6.54 percentage points. In a domain suite spanning seven areas — claim verification, healthcare, cybersecurity, programming, government, law and finance — Decider 1.3 leads TypeSafe’s Jev 1.14 across the board, by up to 12.9 points in finance and 9.3 points in law.

On a sample of one million real API requests, v1.3 agreed with frontier-model judges 93.98% of the time, slightly above v1.2’s 93.68%. Decisions API pricing is $0.02 per million input tokens for v1.3 and $0.01 for Flash. Perplexity notes public leaderboards can be gamed through repeated submissions, and its internal evaluations are all zero-shot.

Key facts

1. Release: Perplexity released Decider 1.3 and Decider Flash 1.3 on Nov 12, succeeding v1.2 and Flash v1.2
2. What it does: Decision models return probability distributions for yes/no, multiple-choice or score questions in a single forward pass, instead of generating text
3. Locate: New feature extracts exact matching text spans with start and end positions; text-only, capped at 128 locate questions per request
4. Size: Decider 1.3 is built on Qwen3.8-27B, about 49 GiB of BF16 weights; Flash is built on Qwen3.5-4B, about 8.5 GiB, distilled with a KL-divergence objective
5. Open: Weights ship under Apache 2.0 on Hugging Face; also on OpenRouter and the Decisions API, with a 262,144-token context
6. Evaluation: 91.6% mean accuracy across 63 classification tasks, 6.54 points above production classifiers; 93.98% agreement with frontier judges on one million real requests