發生什麼事
A decision model reads a scenario and then answers multiple-choice, yes/no or rating questions about it, attaching a probability to each option; sorting incoming customer-service requests is one place this kind of judgment fits. In v1.2, every question in a single request is handled together in one sequence, so a request with several questions costs roughly the same as one with a single question (pplx-decider-v1.2 model page).
The 27B model sits on top of Qwen3.8-27B. Its overall Decision Index 0.3 score climbed from 62.30 in v1.1 to 68.09, and its vision score went from 73.22 to 77.43 (pplx-decider-v1.2 model page). The 4B Flash model is roughly a seventh of that size and was distilled from the 27B model (pplx-decider-flash-v1.2 model page); no other 4B model scores higher on the Decision Index (@perplexitydevs).
01Sources
- @perplexitydevs: Perplexity's official developer account, 10/16 18:42:59 UTC (02:42 on 10/17 Taipei), 73 likes / 8 replies; two sizes of new multimodal decision models with open weights, better text and vision scores, as much as 5x quicker inference, Flash ranked first among 4B models on the Decision Index
- @AravSrinivas: Perplexity CEO, 10/16 19:46 UTC, 86 likes / 20 replies; Decisions API and open weights on Hugging Face, 1 cent and 2 cents per million tokens
- pplx-decider-v1.2 model page: Hugging Face; Qwen3.8-27B, joint requests in a single sequence, Decision Index 0.3 and vision scores, multiple-choice / yes-no / rating questions with probabilities, customer-service routing example, Apache 2.0
- pplx-decider-flash-v1.2 model page: Hugging Face; about one-seventh the size, Qwen3.5-4B, distilled from the 27B model, Apache 2.0