AI news · October 18, 2026
Perplexity ships a small model for typed decisions
Decider Flash V1.2 scores 58.88 on its joint Decision Index with about 9 GiB of BF16 weights; use it for repeatable routing, not open-ended chat.
- Flash parameters
- 4B
- joint Decision Index
- 58.88
- approximate BF16 weights
- 9 GiB
Vercel announced Perplexity's Decider V1.2 models on October 16 for content classification, request routing, and scoring of text, structured records, and images. The API accepts typed questions such as yes/no, choice, or a score against a rubric, and several questions can share one input. The smaller Flash model is aimed at high-volume labeling and filtering; the larger 27B model is for more involved judgments.
The Flash model card reports an overall joint Decision Index of 58.88 and a vision score of 72.53. It uses about 9 GiB for BF16 weights plus working memory, needs Python 3.12 and CUDA, and currently requires a Hugging Face token for the local checkpoint. The useful design choice is explicit answers with probabilities, which lets an application send uncertain cases to a person instead of hiding the decision inside free-form text.
What you can do with it
Write three fixed decisions for one queue, such as damaged, urgent, and route-to-team. Run one question per request on Flash, compare it with rules and your current model, then log agreement, human overrides, latency, and cost. Keep only decisions with a clear review path.
Our take
Typed decisions are easier to score than a chat answer, and the small model may fit high-volume routing. The public scores are model-card results, not proof on your data. Keep the rubric short, save the probability, and make low-confidence results wait for review.
Links Try itDownload model card
Source: Vercel AI Gateway — Made With Models writes the brief; the reporting is theirs.