Perplexity adds span extraction to models that return probabilities

single source· 1 articles · confidence: low · first seen 2026-11-13 16:10 UTC

What this means for you

Nothing to act on yet. Perplexity calls these open-weight but gives no licence, no parameter counts and no download details, so you cannot check the 91.6% against your own routing traffic. The score also carries no evaluation date or harness.

Perplexity released version 1.3 of Decider and Flash, a pair of open-weight models that output probabilities rather than generated text. The update adds Locate, which pulls a specific span out of a document instead of classifying the whole input. Perplexity reports 91.6% mean accuracy across 63 query-classification tasks; no evaluation date or harness is given, so the figure cannot be set against other vendors' numbers. Parameter counts, licence terms and download details were not published. These are routing and labelling models, not chat models.

Key facts

  • ·Perplexity released Decider v1.3 and Flash v1.3, open-weight models that return probabilities rather than text. source
  • ·Version 1.3 adds Locate, which extracts spans from a document rather than classifying the whole input. source
  • ·Perplexity reports 91.6% mean accuracy across 63 query-classification tasks. source
  • ·The report gives no evaluation date, harness, parameter counts, licence terms or download details for the two models. source
  • ·The release was reported on 13 November 2026. source

What the sources say

  • MarkTechPost — Slim release note: version number, new span-extraction mode, one accuracy figure, no licence or weights detail.

Sources

The original reporting. Follow these — they did the work.

← the wire