Perplexity has published a 27-billion-parameter decision model on Hugging Face, and its own scoreboard shows a narrow lead over the rival it is measured against. A decision model does not write an answer. You hand it a message and a list of options, and it returns a probability for each, so your own code decides what happens next. The model card’s example sorts a note about a failing Stripe integration among billing, technical support and sales.
Perplexity reports an overall score of 85.71 percent, against 84.51 for Jev and 74.76 for Qwen3.8-27B, the model it was fine-tuned from. The card says the figures were measured through the Perplexity API, so this is a self-run comparison, not an independent test. The lead is uneven. The new model beats Jev on five of 11 benchmarks, most sharply on RAGTruth (88.80 against 77.27), while Jev wins six, including BBH by about 11 points. On JevBench public hard, the new model even trails its own base.
Running it takes a CUDA GPU with room for roughly 49 GiB of weights. A 1.2 point overall margin is thin, so any team routing tickets or screening documents should test it on its own data.
Reported from Perplexity’s model card on Hugging Face. The page carries no publication date.
