Perplexity pplx-decider-v1-27b: Open-Weight Decisions

Publisher, base and model size

The repository belongs to the verified Perplexity Hugging Face organization. Its exact ID is perplexity-ai/pplx-decider-v1-27b; the model card names Qwen/Qwen3.8-27B as its base and Apache-2.0 as its license. Hugging Face reports 26,085,330,160 parameters (about 26.1B); the product name retains “27b.” Model card.

Typed output, including image decisions

The repository’s inference wrapper accepts text state and optional images. Its shared schema defines Choice, Noul and Score, with probability fields rather than a prose-only response. The README demonstrates Choice and Noul; Score is confirmed in the published schema. Inference wrapper · Typed output definitions.

Self-host access and resource requirements

The official model card specifies Python 3.12+ and a CUDA GPU with room for approximately 49 GiB of weights plus working memory. That is not a total VRAM guarantee: inputs and runtime overhead need additional capacity. The supplied example downloads weights and executes repository code; inspect and pin it before use.

The card says its results were measured through the Perplexity API. It does not establish a public endpoint, price or general API entitlement. This profile confirms the open-weight route, not a generally available hosted service.

How to read the eleven-benchmark table

Perplexity reports an 85.71% overall aggregate across 11 benchmarks in its model card. This is a publisher-reported result, not an AI Tool Finder independent winner; per-task outcomes vary. We did not reproduce the evaluation. Read the full benchmark table.

Evaluate your own labels and failure costs. Keep uncertain decisions reviewable, and distinguish valid response shape from correct judgment.

Official sources

• Official Perplexity model card
• Official inference example
• Choice / Score / Noul types

Choose your next step