Perplexity's Decider v1.3 Doesn't Generate Answers — It Only Outputs Probabilities
Perplexity has released Decider v1.3 and Flash v1.3, two open-weight decision models that output probabilities directly instead of generating natural text. The v1.3 update adds a new Locate feature for span extraction, achieving an average accuracy of 91.6% across 63 query classification tasks.
Perplexity, the AI search firm, has rolled out two open-weight models: Decider v1.3 and Flash v1.3. They do not produce paragraphs of natural text, and output probability values directly. You can understand them as purpose-built models for tagging queries and running deterministic judgment, which are perfect for use cases such as classification, routing, and information extraction.
The core update for v1.3 is the new Locate function. It identifies the exact span that contains the target answer in the input text, returns its position, and generates no extra explanation. As stated in the official report, the model reaches an average accuracy of 91.6% on 63 different query classification tasks. This metric proves its stable performance on these specific tasks, but the result cannot be directly extrapolated to open-domain question answering — classification tasks and generative tasks are fundamentally different.
Open-weight means full model parameters are available for public download, allowing development teams to deploy, fine-tune and evaluate the models in their own isolated environments. For AI application developers, the value of this type of model lies in cost efficiency and controllability: calling a full large language model just for simple judgment tasks is both expensive and slow. The standard industry engineering practice is to use a small lightweight judgment model for pre-routing first, and then pass relevant requests to full generative models later.
That said, probability outputs also bring extra interpretability costs. Setting appropriate classification thresholds, calibrating confidence scores, and building proper error fallback mechanisms are all issues that need to be handled during real-world implementation. By integrating Locate into this release, Perplexity is clearly aiming to put both judgment and span location capabilities into the same model layer. Its actual performance remains to be verified in real production tasks. Source: MarkTechPost, by Asif Razzaq. Original Article Link.
发布时间: 2026-11-14 00:10