Perplexity Decider Flash v1.2 Goes Open

This is a compelling shift from chat-centric models toward compact, typed decision engines that application code can actually trust and control.

• –Distilled from Perplexity’s 27B decision model, Flash retains strong performance at roughly one-seventh the size.
• –Its separate decision head produces probabilities directly, avoiding fragile parsing of natural-language responses.
• –Apache 2.0 weights enable self-hosting, fine-tuning, and deployment in latency-sensitive pipelines.
• –Developers should benchmark question batching carefully: the model recommends one question per request except for tightly coupled multi-label decisions.
• –The custom noncausal runtime and roughly 9 GiB of BF16 weights make this less plug-and-play than a conventional Transformers model.