Perplexity open source 27B decision-making model, the official evaluation comprehensive score slightly exceeds Jev

Dongcha Beating AI News, Perplexity launches Decisions API and open source multi-modal decision model pplx-decider-v1-27b. It does not generate ordinary text, but directly returns the probability distribution in the answer given by the developer, which can be used for classification, request routing and Agent next action selection. API pricing is $0.04 per 1 million input tokens, and outputs are free.

Among the 11 items and a total of 7210 sample reviews published by Perplexity, the comprehensive accuracy rate of pplx-decider-v1-27b is 85.71%, which is higher than Jev's 84.51% and Qwen3.8-27B's 74.76%. But it's not leading across the board. Jev scored higher on 6 of the 11 tests, with Perplexity leading FinancialPhraseBank, RAGTruth, TabFact, and Circa. These scores are from Perplexity's own self-assessments, not independent third-party reviews.

This type of model specializes in handling high-frequency "small judgments" in Agent. For example, who should assign the customer service ticket to, whether the tool should be called, and which process to take next. Traditional LLM usually generates a piece of text and then parses the answer, while the decision model directly returns preset options and probabilities. AWS, Cloudflare, etc. have recently launched similar models.