Perplexity Decider v1.2 27B

pplx-decider-v1.2 succeeds pplx-decider-v1.1-27b on the same Qwen3.8-27B backbone. Two things changed: the request format and the training data. v1.1 answered one question per forward pass by reading out the last token; v1.2 reads the whole request in one sequence, with each question getting its own readout at a marker token placed after that question's options. Several questions about the same state can therefore inform each other, and the checkpoint handles text and images up to 1.6 MP. The checkpoint combines several training runs through model souping (weight averaging), chosen as the best compromise across evaluations rather than on a single benchmark. On the public Decision Index 0.3 the model scores 68.09 with joint requests and 67.75 with one question per request, against 62.30 for v1.1. On the Vision Decision Index 0.3 it scores 77.43 against 73.22 for v1.1. Running it locally needs Python 3.12+ and a CUDA GPU with room for about 52 GiB of BF16 weights plus working memory; the Hugging Face repository is private and needs a token with access.