Perplexity: pplx-decider-flash-v1.2

Perplexity: Decider Flash V1.2 is a compact, open-weight decision model distilled from Perplexity's larger decision models and built on the Qwen3.5-4B backbone with 4 billion parameters. Rather than generating explanatory text, it answers named yes/no questions with probabilities, choice questions with a probability per option and a most-likely selection, and rubric-score questions with a probability per level plus an expected score. It accepts text, JSON, image, and PDF input, supports structured output, and shares the same API contract as the full Decider V1.2 model. A 262K token context window is notable for a model of this size, and it processes a whole request in a single pass so multiple questions about the same input can be answered together. It was trained with model souping, which averages weights across training runs to improve robustness. The model is intended for high-volume, decision-style workloads where software compares a numeric answer against a threshold: classification, routing support tickets to the right teams, filtering, labeling messages, content moderation, sentiment scoring, rubric grading, and structured data extraction or attribute verification from images. Because it outputs probabilities, it fits pipelines that need a confidence value rather than free-form commentary, and it can be run locally on compatible hardware at about 13 GB of VRAM at 4-bit quantization or on-device through the Decider for Mac app, which requires 5.8 GB of storage and 16 GB of memory. It is available through Perplexity's API, the Opper gateway, or locally, and is positioned as the cost-efficient option in Perplexity's decision model lineup. Users should keep several constraints in mind. The model is designed for one question per request; multi-question requests process and bill the state again for each question. No explicit reasoning mode is documented. Training used sequences up to about 32K tokens, with longer contexts spot-checked rather than broadly validated, so real-world accuracy on very long inputs may vary. Perplexity recommends validating accuracy on your own labeled examples, especially for nuanced judgments, and independent testing of the larger 27B Decider model showed weaknesses in contextual visual judgment tasks.