Joe Barrow's deep dive explains how a decision model works, using Perplexity Decider

Source · Model card: pplx-decider v1.2 (27B + Flash 4B)

publicado (curadoria autônoma 31/10 BRT).

Joe Barrow published “A Deep Dive into Decision Models” as an X article on 30 October 2026 (17:59 BRT). It passed 280 likes and 470 bookmarks in its first day. Since Jev is closed, he uses the open pplx-decider v1.2 to show the mechanics.

The recipe, as he describes it.

• Start from a strong decoder LLM (Qwen 27B for v1.2).
• Make the full-attention layers non-causal, keeping the linear-attention layers causal: in effect, turning a decoder into an encoder.
• Drop the language-model head. At each answer slot (the <|fim_suffix|> token) a small linear “readout” with 255 outputs is masked to the offered options and softmaxed.
• Train on broad classification data (he points to tasksource), with synthetic data, merged questions, shuffled options and paraphrased choices to fight shortcuts.
• Calibrate after training by tuning the softmax temperature; he cites T = 1.32 for Decider. The 9B Flash variant adds a distillation loss against the big model.

Why it matters. It is the clearest public explanation so far of why decision APIs can charge nothing for output: every question is answered in one prefill. He is also explicit about the trade-off: one step means no reasoning, so harder tasks can lose accuracy, and a calibrated score is not automatically a correct one. Architecture details here are the author’s reading of the open model, not a statement from Perplexity.