Perplexity releases open-weight Decider v1.3 and Flash v1.3 classification models

Perplexity released pplx-decider-v1.3 (26B) and pplx-decider-flash-v1.3 (5B) as open-weight (Apache 2.0) decision models that return probability distributions over answer options rather than generating text.
The v1.3 model achieves an average score of 69.0 and Flash 63.6 on 17 out-of-domain NER, PII, and extractive-QA tasks, compared to 43.3 for Liquid d1 and 33.6 for GLiNER large v2.5.
Both models support text and image inputs and are available via OpenRouter and Perplexity's API, with BF16 weights requiring approximately 49 GiB for the 26B model and 8.5 GiB for the 5B model.
A new locate question type extracts verbatim spans with character offsets using a separate LoRA adapter, which remains unmerged to prevent regression in classification performance.

Technical Specifications

Fine-tuned from Qwen3.8-27B (v1.3) and Qwen3.5-4B (Flash), with the Flash model distilled using a KL objective on the teacher's distributions.
Backbone full-attention layers are bidirectional, requiring the pplx_decider package for inference as plain causal inference does not reproduce outputs.
Training sequences reached up to about 30k tokens, and the models support yes/no, multiple choice (up to 255 options), ordered scores, and locate question types.

Multilingual classification across 8 languages reached 88.16% for v1.3 and 86.50% for Flash, exceeding TypeSafe's Jev 1.14 at 85.41%.
Reranking performance hit 83.99 nDCG@10 across 72 retrieval tasks, a 1.81 point improvement over Qwen3-Reranker-8B.
Artificial Analysis Decision Index scores are roughly 70 for v1.3 at $68 per 1M tasks and roughly 64 for Flash at $34 per 1M tasks.