Perplexity open-sources Decider v1.2 and Flash v1.2

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

publicado (OK do Fabricio 17/10 BRT).

On 15-16 October 2026 Perplexity put two new decision models on Hugging Face and in its Decisions API: pplx-decider-v1.2 (27B, Qwen3.8) and pplx-decider-flash-v1.2 (4B, Qwen3.5, distilled from the 27B models). The Perplexity Devs account announced them on 16 Oct; CEO Aravind Srinivas called the 27B “#2 on HF Decision Index overall - for now” and the 4B the top 4B model. Weights are Apache-2.0.

What is new. Requests are now joint: the state and all questions go into one sequence, with a readout after each question, so several questions about one state cost about as much as one. Perplexity says inference is “up to 5x faster” for that reason (not reproduced). Its own table shows that grouping helps when questions are facets of one decision and hurts when they are competing candidates, and it recommends one candidate per request when in doubt.

Independent result. The official Decision Index 0.3.2 board (data of 16 Oct 19:00 UTC) lists the 27B at Full 65.55, #4 (Jev 1.13: 60.11) and Flash at 54.81, #47. The card’s own public-suite figures are 68.09 (27B) and 58.88 / 60.05 (Flash, joint / one question per request). Perplexity states it added the benchmarks’ official training splits (decontaminated) to the data, so the Full score is the number to compare.

Prices and routing. The Decisions API charges $0.02 per million input tokens for v1.2 and $0.01 for Flash, output free, 10 requests per second per organization. Requests that name the old pplx-decider-v1.1-27b or pplx-decider-v1-27b now run on v1.2. The Hugging Face cards still say the repository is private; it is public.

Catalog entry: pplx-decider v1.2 and Flash (27B scored 6.3 / 8.1 / 18). The v1.1 page is kept as history: pplx-decider v1.1.