Perplexity releases Decider v1.3 open-weight models with a span-locating task
Perplexity's Decider v1.3 adds a locate task that returns span offsets, and the company says it beats its production classifiers by 6.54 points.
Perplexity has released version 1.3 of its Decider open-weight decision and extraction models. The update adds a "locate" task that returns span offsets, meaning the start and end positions of a matched piece of text within the input. The company says the task covers named entity recognition, which picks out names of people, places, and organizations; PII detection, which flags personal information; and extractive question answering, where the answer is lifted directly from a passage. A label alone tells a developer what was found. Offsets tell them where it sits in the source text.
Aravind Srinivas, who announced the release, says it is 6.54 points more accurate than Perplexity's production classifiers across 63 tasks. That accuracy figure is the company's own claim, measured against its own production classifiers. The models are free on OpenRouter for one week, according to the same post.
What changed
Decider v1.3 adds a locate task that returns span offsets for named entity recognition, PII detection and extractive question answering.
Type: New capability · Applies to: Decider v1.3
Why it matters
Teams running extraction or PII redaction could test a free, span-level model against their current classifiers.
What you can do
Run your own PII and extraction evaluation set against it during the free OpenRouter week.
Original source
We have no link to the original source for this story yet.
Where we saw it
Latent Space
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works