Perplexity Ships Decider v1.3 Classifier, 91.6% on 63 Tasks

Perplexity on November 12 released Decider v1.3 and a smaller distilled version called Decider Flash v1.3, publishing both under Apache license on Hugging Face. 'We are releasing Decider v1.3 and Decider Flash v1.3, the successors to our earlier decision models: v1, v1.1, v1.2, and Flash v1.2,' the company writes in a blog post on its hub.

The larger model posts 91.6% mean accuracy across 63 query-classification tasks, which Perplexity says exceeds its own production classifiers by 6.54 percentage points. Across eight languages (German, Spanish, French, Italian, Russian, Japanese, Korean and Chinese), the mean is 88.16%. On a separate 72-task retrieval and reranking bank it scores 83.99 nDCG@10, and in seven domain-specific sweeps the post claims leads over a rival called Jev 1.14 of up to 12.9 points in finance and 9.3 in law.

A 'decision model,' as Perplexity defines it, is a general-purpose classifier. 'You give it an input and a set of questions each with its possible answers,' the post says, covering three answer shapes: yes/no, multiple choice, and ordered score. The company starts from generative Qwen weights and turns the attention bidirectional, arguing that 'making the backbone's full-attention layers bidirectional improves performance, so we adopt this design to let each token incorporate context from the entire input.'

Pricing is the pitch. Running a million production requests costs $26.47 on Decider v1.3 and $13.24 on Flash. 'Both models are available on Hugging Face under Apache license, through OpenRouter, and the Perplexity Decisions API,' the authors write. Across 2.3 million questions pulled from a million real requests, Perplexity reports a 93.98% agreement rate with its human judges.

It lands in a week we've been cataloguing a steady flow of Apache-licensed open-source releases.