Perplexity Releases Decider v1.3 Decision Models with New Locate Feature

Perplexity AI released Decider v1.3 and Decider Flash v1.3, open-source decision models available on Hugging Face (Apache license), OpenRouter, and the Perplexity Decisions API.
The models achieve 91.6% average accuracy across 63 internal classification tasks, surpassing previous production classifiers by 6.54 percentage points and Jev 1.14.
A new 'locate' feature enables named entity recognition, PII identification, and extractive QA by returning matching text with start/end character positions.
Architecture uses bidirectional attention and an 'Isolated questions' pattern where questions attend only to the state and their own tokens, ensuring result stability.

Performance Benchmarks

Decider v1.3 scores 83.99 average nDCG@10 on 72 search tasks, beating Qwen3-Reranker-8B by 1.81 points and Jev 1.14 by 2.24 points.
Multilingual evaluation across 8 languages and 46,484 examples shows Decider v1.3 averaging 88.16% accuracy, outperforming Jev 1.14 (85.41%) in all languages.
On 17 out-of-domain locate tasks, Decider v1.3 scores 69.0, significantly higher than Liquid d1 (43.3) and GLiNER large v2.5 (33.6).

Analysis of 1 million production requests over 18 days shows Decider v1.3 has a 93.98% agreement rate with frontier LLM judges, compared to 93.35% for Jev 1.14.
Label sensitivity tests indicate Decider v1.3 changes answers in only 0.27% of cases during label reordering, whereas Jev 1.14 changes 2.69% to 4.84%.
Internal use cases include security policy improvement (processing 1 billion tokens in one day), RL post-training monitoring, and web crawling quality filtering.