New decision model released with an extraction feature
Perplexity on Nov. 12 (3:33:50 a.m. Nov. 13 Japan time) released "pplx-decider-v1.3," a new version of its decision-only AI model with open weights, along with a smaller version, "pplx-decider-flash-v1.3." Its official developer account on X, @perplexitydevs, announced the release. v1.3 adds "Locate," a new feature that extracts matching passages from text based on type or instructions. The models are offered through the company's own "Decisions API," the AI model hosting site Hugging Face, and OpenRouter, which provides access to many AI models through a single gateway.
We're rolling out pplx-decider-v1.3 and pplx-decider-flash-v1.3, the latest version of our open-weight decision models. v1.3 adds locate, a new feature that extracts spans based on type or instructions. Now available in our Decisions API, @huggingface, and @OpenRouter. https://t.co/Qhf4LcYA4s
According to the Hugging Face model page, v1.3 is the successor to "pplx-decider-v1.2" and is a classification model that handles both text and images. Instead of generating text, it answers questions in three forms: choosing one option from a list, giving the probability of yes or no, and assigning a score on a scale. The company said it also improved performance and robustness on decision tasks. On Oct. 22, Perplexity released "Decider for Mac," an app that runs the previous small model, "pplx-decider-flash-v1.2," on Macs.
Locate returns phrases with positions and scores
According to the model page, Locate extracts matching words or phrases verbatim from the input text. The target is set by the name of the question; naming it "person," for example, extracts people's names. Users can also narrow the target by adding an instruction such as the person who signed the contract. Answers come back as sets of the extracted text, its start and end positions in the text, and a score showing confidence. Locate questions and the existing decision questions can be sent together in a single request.
Locate works only on text and cannot be used in requests that include images. A single request can carry up to 128 questions. The model page says requests that mix decision and Locate questions are processed separately for each type, so they take longer than requests with only one type.
About 26.1 billion parameters for the large model, 4.5 billion for the small one
pplx-decider-v1.3 is built on Qwen3.8-27B from China's Alibaba. File information on its Hugging Face page puts the parameter count at about 26.1 billion. The page says running it on your own hardware requires a GPU with room for about 49 GiB of weights in BF16 format plus working memory.
The smaller pplx-decider-flash-v1.3 is based on Qwen3.5-4B and has about 4.5 billion parameters. It was built by transferring knowledge from the large model, and it shares the large model's request format, API, and support for images and Locate. The model page says it suits uses where speed or cost matters more than the last small margin of accuracy.
Hugging Face pages list the license for both models as "Apache 2.0." The post and the model pages do not state pricing for v1.3 on the Decisions API.
