Perplexity Decider v1.2 was the easiest release to miss on October 17, 2026: Perplexity updated the decision model's card on Hugging Face, Vercel listed Decider V1.2 27B and Decider Flash V1.2 4B on AI Gateway at the same time, and Perplexity CEO Aravind Srinivas posted the same day that the Flash version can run on-device on an M4 Mac with 16GB of memory. It takes a completely different road from chatbots: it writes no replies and generates no code. It only answers pre-framed judgment questions, with probabilities.
It returns only three kinds of answers, each with a probability
A decision model is used by handing it material (called state in the model card: text, structured records, or images) plus one or more typed questions. Boolean questions return the probability of yes, from 0 to 1. Choice questions assign a probability to each of your options and pick the most likely one. Score questions rate against an ordered rubric you define, level by level. Your application then decides what to do with those numbers: which team gets a ticket, whether a product photo shows damage, or whether an uncertain call goes to a human. Because nothing is generated word by word, these models change the arithmetic for high-volume classification, routing and labeling.
What v1.2 changes: related questions can be asked in one request
This version is fine-tuned from Qwen3.8-27B and accepts images up to about 1.6 megapixels in its input. The new trick is putting one piece of material and several questions into a single request, each question with its own readout marker, so related questions share context at roughly the compute cost of one. On Perplexity's own Decision Index 0.3, the joint-request configuration scores 68.09 overall, up from 62.30 for v1.1; its Vision Decision Index rose from 73.22 to 77.43. Two discounts apply: Perplexity maintains and runs this benchmark itself, and this version's training data now includes the official training splits of several of its benchmarks, so how much of the gain is better judgment versus familiarity with the exam needs third-party testing to separate. Perplexity itself warns that scoring independent candidates should be done in separate requests, because grouping them lets them influence each other.
The Flash version wants to move judgment on-device
The smaller Flash sibling has 4B parameters and targets high-throughput labeling, filtering and routing. Vercel's listing likewise reserves the 27B model for nuanced judgments and detailed rubrics, and Flash for high-volume tasks with clear criteria. If Srinivas's on-device claim holds up, part of classification and first-pass screening could stop sending material to the cloud — a real option for privacy-sensitive, high-volume work. For now it is a public statement from the CEO with no published on-device speed or accuracy numbers, so do not design an architecture around it yet.
Self-hosting the 27B version also has a price to check first: BF16 weights need about 52 GiB of GPU memory plus working memory, with Python 3.12 or later and a CUDA GPU. It is not a small component to drop into an existing service. The steadier reading is this: within weeks of Jev appearing, the decision-model lane has filled with Perplexity, Cloudflare, OpenAI and others. The point of v1.2 is not the score itself, but that judgment is being split into its own cheap, potentially on-device step — and anyone whose business runs on large numbers of yes-or-no questions should start testing it.
