Perplexity open-sources pplx-decider-v1.1-27b: Decision Index rises from 56.4 to 61.56, beating Jev by 3.6 points

Perplexity's pplx-decider-v1.1-27b is an Apache-2.0 open-weights decision model fine-tuned from Qwen3.8-27B. Instead of generating text it outputs calibrated probabilities over the candidates you define, via a BF16 [255, 5120] readout head rather than a full-vocab lm_head. Versus v1, it lifts the causal mask on full-attention layers and trains on more data (including tasksource), raising the overall Decision Index from 56.4 to 61.56, ahead of Jev at 57.9 (Language 69.45, Retrieval 61.26, Tools 78.88, Knowledge 48.18; all vendor-reported). Self-hosting needs Python 3.12+ and a CUDA GPU for ~49 GiB of weights, using the bundled DecisionModel code to reproduce evaluated behavior. Alternatively call POST https://api.perplexity.ai/v1/decisions with a state and named noul/choice/score questions. Good fit for agent intent routing, RAG relevance checks, and rubric grading.