What changed
Perplexity released Decider 1.3 and Decider Flash 1.3 for probabilistic classification, scoring, ranking and multimodal decisions. The new locate mode extracts exact text spans with offsets, and Perplexity publishes Apache-2.0 model artifacts alongside access through OpenRouter and the Perplexity Decisions API.
Why it matters
A bounded decision model can replace free-form generation for routing, moderation, retrieval, extraction and evaluation while returning probabilities. The smaller Flash variant creates a lower-resource option, but thresholds, calibration and access conditions remain workload-specific.
What to do
Build a labeled holdout set with the cost of false positives and false negatives recorded, compare both variants with the current classifier, tune thresholds and calibration, and validate exact-span offsets, latency, memory and provider costs before routing production traffic.
Caveat
Perplexity's benchmark, cost and production-log results are vendor measurements, including LLM-judged references and workload-specific cost accounting. Although the artifacts are labeled Apache-2.0, the current Hugging Face pages require an authenticated token with repository access; inputs are not truncated and the 27B BF16 checkpoint needs roughly 49 GiB plus working memory.