Perplexity ships Decider v1.2, a faster multimodal decision model
A joint request format lets Decider answer several questions about the same input in one pass, with up to 2.74x higher throughput and open Apache 2.0 weights.
Perplexity released pplx-decider-v1.2 on 16 October, retiring pplx-decider-v1.1-27b, and introduced a smaller sibling, pplx-decider-flash-v1.2. Both are decision models: built to answer classification-style questions about a piece of text or an image rather than hold a conversation.
The headline change is a joint request format: several questions about the same input now run in a single forward pass instead of one call each, so the questions can inform one another. Both models take images as well as text, up to 1.6 megapixels, and Perplexity says a torch.compile-based fast path gets up to 2.74x higher throughput on decision requests and 2.49x on reranking versus standard inference. On Perplexity’s own Decision Index 0.3 benchmark, the overall score rose from 62.30 to 68.09, and the vision score from 73.22 to 77.43. Weights are released under Apache 2.0 on Hugging Face, with API pricing at $0.02 per million input tokens for the 26B model and $0.01 for Flash; output tokens are free, and requests still naming the old v1.1 model now resolve to v1.2.
Why it matters: if you’re using an LLM purely to classify, rank or route rather than to chat, a dedicated decision model that answers several questions in one pass is cheaper and faster than running a general chat model per question, and the open weights mean you can self-host instead of paying per call.
The caveat: the benchmark gains are Perplexity’s own Decision Index, not an independent evaluation.