Perplexity refreshes its open decision models: v1.2 ships in two sizes, and gets its first third-party grade
From the AI digest for 2026-10-17
What happened: On October 16, 2026 at ~3:46 PM ET, Arav Srinivas announced the next generation of Perplexity's open-weight decision models: pplx-decider-flash-v1.2 (4B parameters, 1 cent per million tokens, the top 4B model on the HF Decision Index) and pplx-decider-v1.2 (27B parameters, 2 cents per million tokens, #2 on the HF Decision Index overall). Perplexity's developer account followed with the detail that text and vision scores improved and inference is up to 5x faster because all questions in a request run in one pass. Clem amplified at ~4:41 PM ET, noting the weights are on Hugging Face. On October 17 at ~4:10 PM ET, Arav then amplified third-party testing of Flash: "Perplexity Decider Flash is pretty good and can run super fast and on-device on your M4 Mac 16GB" — pointing to Wolfram Ravenwolf's account of tests in which Decider Flash "matched Wald's accuracy at nearly twice the speed on a 16 GB Mac mini" across "synthetic agent decisions, real email threads, prompt injections," and is Ravenwolf's new pick for email-completion checks.
What's genuinely new: A development of the decision-models thread (the TypeSafe AI's Jev storyline), which already tracks the earlier pplx-decider generations, Strands Decider, and the still-missing independent head-to-head. New here: a version bump from v1.1 to v1.2 with concrete parameter counts and per-million-token USD pricing for both sizes, actual open weights published on Hugging Face, and a measurable inference-speedup claim. The October 17 layer adds the record's first third-party performance claim for the family — on-device operation on a 16 GB consumer Mac with a claimed accuracy-parity-at-roughly-2x-speed result — though it reaches the record through the vendor's own amplification, and the comparator named only as "Wald" (model identity not identified in capture) is unverified. Unlike the Roadster beat, this one is shipped — weights, API availability, and now a named consumer-hardware deployment path are all claimed live, though the benchmarks remain vendor- or vendor-relayed-self-reported.
Watch: An independent run of v1.2 against Jev, Clef, and the OpenAI Decisions API — the comparison the thread has never captured; the published Ravenwolf methodology and what "Wald" actually is; whether Jev, Clef, or the Decisions API get the same on-device treatment; adoption of the 5x single-pass inference framing elsewhere; whether the "#2 overall, will improve rapidly" HF Decision Index position holds.
Links: Arav Srinivas announcement (X) · Clem RT on Hugging Face weights (X) · Arav on Decider Flash on-device (X)