Perplexity releases pplx-decider-v1.2 for joint text and image decisions

Perplexity AI detailed on Hugging Face pplx-decider-v1.2, a decision model built on the Qwen3.8-27B backbone. The checkpoint handles text and images up to 1.6 megapixels.

The model processes a state and multiple questions in one sequence, with a separate readout marker for each question. This lets related questions share context while using about the same computation as a single question. Perplexity recommends separate requests when scoring independent candidates because grouped candidates can affect one another.

On Decision Index 0.3, a benchmark suite Perplexity publishes and ran itself, the joint request configuration scored 68.09 overall, compared with 62.30 for pplx-decider-v1.1-27b; Perplexity says the training data now includes the official training splits of several of the suite's benchmarks. Its Vision Decision Index score rose from 73.22 to 77.43. Running the model requires Python 3.12 or later, a CUDA GPU, and about 52 GiB for BF16 weights, plus working memory. The Hugging Face repository is private and requires an authorized token.