Deyron Labs graphic: Perplexity Decider v1.3: open-weight decision model; Yes/no, choice and score questions with probabilities; Weights on Hugging Face under Apache, per Perplexity; 91.6% mean accuracy on 63 tasks (Perplexity's own); Cost for 1M requests: $26.47 (Flash $13.24)

What happened

On 12 November 2026 Perplexity Research published “Perplexity Decider v1.3: Training, eval, and use cases”. Decider is a general-purpose classifier: for yes/no, multiple-choice and score questions it returns a probability distribution over the answers. Version 1.3 also adds a “locate” feature that extracts exact text spans, such as named entities, personal data or answers to extractive questions. A smaller Decider Flash v1.3 is distilled from it.

Key details

All figures are Perplexity’s own, from its post read on 14 November 2026. We have not run the models.

  • Availability: Perplexity says the weights are downloadable from Hugging Face under an Apache license, and that the models are also offered through OpenRouter and the Perplexity Decisions API.
  • Base models (as stated): Decider was trained from Qwen3.8-27B and Flash was distilled to Qwen3.5-4B. Perplexity says the design was inspired by TypeSafe AI’s Jev.
  • Accuracy claims: 91.6% mean accuracy over 63 classification tasks (6.54 points above Perplexity’s production classifiers); 83.99 mean nDCG@10 over 72 retrieval tasks (1.81 above Qwen3-Reranker-8B); 88.16% mean multilingual accuracy; 93.98% agreement with LLM judges on 1 million real requests (93.35% for Jev 1.14).
  • Locate: an average score of 69.0 versus 43.3 for Liquid’s d1, per Perplexity.
  • Cost comparison for 1 million requests: $26.47 for Decider v1.3, $13.24 for Flash, $68.10 for Jev 1.14, $312.21 for GLM-5.3-Flash and $349.48 for DeepSeek V4.1 Flash. Perplexity notes these are workload-specific.
  • Not stated: a per-token API price, and independent evaluations.

Why it matters

Decision models are becoming a standard category: small, cheap models for routing, grading and checking rather than for chat. Google, Microsoft and OpenAI have launched their own versions and Liquid released open d1 models; all but the last are covered on this site. Perplexity’s entry is notable for open weights under a permissive license, which lets developers run it themselves.

The comparisons are against Perplexity’s own baselines and chosen tasks, so a test on your own labelled data is the fair check. The license and exact weights listing on Hugging Face should be confirmed before use in a product.

Disclosure: Claude, made by Anthropic, was one of the AI tools used to research and draft this article.

Sources

  1. Perplexity Decider v1.3: Training, eval, and use cases Primary source , Perplexity Research, 12 November 2026