The search company’s router, released to everyone. Perplexity Labs shipped pplx-decider-v1 on October 1, 2026 and a v1.1 refresh on October 5: a 26.1B-parameter text-classification model, Apache 2.0, on the Qwen3.5 27B backbone the rest of the week’s decision models also tune. The contract is the category’s: text in, one label out of a fixed set, confidence number attached, one forward pass, no generated tokens. GGUF builds from the community quantizers (mradermacher covers 12B-class siblings of the same wave) appeared within days.
What Perplexity needs it for tells you where decision models are going. A search engine routes every query: web search vs math vs translation vs shopping vs a clarifying question, in milliseconds, on every request, before the expensive model ever loads. Decider is that router open-sourced - the internal traffic-shape classifier became the product. For local operators the same play applies at home: query routing for a local agent stack, intent gating for voice pipelines, ticket triage for a helpdesk, all on a card that fits in 16GB in 4-bit.
Where it sits in the wave. The October decision-model flood (Clef by Cloudflare, GEV by autotrust, Torchcast 12B) landed within seven days; Decider’s v1-to-v1.1 turnaround inside the same week signals fast iteration, and the Apache license keeps it forkable. Unlike Clef it is text-only - screenshots stay with the vision-tuned members of the family.