Perplexity Releases Decider v1.3 and Decider Flash v1.3 Classification Models
Perplexity has released Decider v1.3 (26B) and Decider Flash v1.3 (5B) as open-weight Apache 2.0 classification models that output probability distributions instead of generating text token by token. Version 1.3 introduces a new 'locate' question type to extract verbatim spans with character offsets for tasks like named entity recognition and PII detection. These decision models provide a cost-effective and highly accurate alternative for classification, retrieval reranking, and extraction tasks without the overhead of traditional autoregressive text generation. Releasing them under an open-weight Apache 2.0 license significantly empowers the open-source AI community to build specialized NLP pipelines. The models are fine-tuned from Qwen architectures with bidirectional attention layers, requiring a dedicated package rather than standard causal inference. While pplx-decider-v1.3 needs about 49 GiB of memory for BF16 weights, the smaller Flash variant uses about 8.5 GiB and was distilled using a KL divergence objective.
## BACKGROUND
Open-weight models provide publicly released parameters like weights and biases that users can download, run, and fine-tune depending on their license. Knowledge distillation is a compression technique where a smaller student model learns to mimic a larger teacher model, often by minimizing Kullback-Leibler (KL) divergence to match output probability distributions.