Perplexity quietly enters the decision-model race, and an independent benchmark already ranks it first

A new entrant has surfaced in the small but increasingly crowded field of open-weight "decision models" — compact systems that return a calibrated probability for a structured question instead of generating free-form text. This week, Ollama's tracking began showing pplx-decider, a model published by Perplexity, appearing among newly surfaced listings with essentially no download history. Within days it had been folded into an independent benchmark that ranks it above several of the category's more established names.

The decision-model category has filled out quickly over the past several weeks, with Typesafe's Jev, Cloudflare's open-sourced Clef, and autotrust's JEV/GEV family all taking turns as reference points for what a "typed" classifier-style model should look like: feed it a state — text, JSON, or an image — plus a set of typed questions, and it returns calibrated probabilities across the possible answers in a single forward pass, with no generated tokens and no sampling variance to manage. Perplexity's entry follows the same basic contract, but it is notable less for the architecture than for who is publishing it. Perplexity is primarily known for its consumer search product and API-hosted sonar models, not for releasing specialized open-weight infrastructure components, which makes pplx-decider a signal that the decision-model pattern is being treated as common middleware worth shipping even by companies whose core business sits elsewhere.

What the model actually does

According to its Hugging Face model card, pplx-decider-v1-27b is a decision model fine-tuned from Qwen3.8-27B, carrying roughly 26 billion parameters in BF16 and tagged for multimodal text classification with image support. It exposes a Decider class with two primary question types: "choice," which selects one labeled option from a set (routing a support ticket to billing, technical support, or sales, for example) and returns calibrated probabilities for each; and "noul," which answers a yes/no question with a probability rather than a hard label (flagging whether a message expresses urgency). The Ollama listing adds a third mode, "score," for probability-weighted ratings across an ordered scale. Running the model locally requires a CUDA GPU with roughly 49 GiB of memory for the weights alone, putting it well outside laptop-class hardware — a contrast with Liquid AI's much smaller 3B d1 model, which targets edge and on-device deployment instead.

On its own 11-benchmark evaluation set, measured through the Perplexity API, the card reports an overall score of 85.71% for pplx-decider against 84.51% for Jev and 74.76% for the unmodified Qwen3.8-27B base model. The comparison is not uniformly favorable: Jev leads on several individual benchmarks, including a 94.27% to 82.80% gap on BBH, while pplx-decider pulls ahead on tasks like RAGTruth and TabFact. The card also discloses that its training pipeline was built with Bespoke Labs' Nimble data tooling, using public human-labeled datasets while explicitly withholding benchmark subsets from training — a detail aimed at pre-empting the contamination concerns that dog most self-reported model comparisons.

An outside ranking, with caveats

The more interesting corroboration comes from outside Perplexity entirely. JevBench, a GitHub Pages benchmark describing itself as an "open benchmark for typed decision models," now scores 17 decision models — Jev, Clef, d1, and pplx-decider among them — across 25 datasets, with results shown as a per-dataset heatmap and an aggregate index. As of this week, JevBench lists pplx-decider v1.1 — a newer release than the v1 build currently mirrored on Ollama — as its top-ranked model, with an index score of 78.0. The benchmark's results drew enough attention to generate a Hacker News submission on October 12, titled "JevBench: Benchmarking decision models (Clef, pplx-decider and 15 more)," though the thread itself drew only a handful of points and two comments, underscoring how early-stage attention to this niche still is.

That independence should be read with some caution. JevBench's page does not identify who operates it beyond the GitHub account hosting it, and its own documentation notes that scales differ across datasets, making cross-dataset comparisons only meaningful within a shared metric — a limitation that applies to the index ranking as much as to any single score. It is a useful outside data point rather than a settled verdict, and it benchmarks a slightly newer version of the model than the one currently distributed through Ollama, which complicates any clean before-and-after comparison.

On raw adoption, pplx-decider remains minuscule next to its rivals. Ollama's tracking shows the model entering its popular list at rank seven with effectively no pull history, and it currently sits at 39 downloads against a 17GB footprint for the 27B build. Liquid AI's d1, by comparison, has already accumulated more than 10,000 pulls on the same platform, and Perplexity's own Hugging Face repository shows a more modest but non-trivial 2,593 downloads in the past month and 122 likes — suggesting the Hugging Face audience discovered the model well before Ollama users did. The gap illustrates how uneven distribution can be even for a model that already leads an independent leaderboard: topping a niche benchmark and getting picked up by the broader self-hosting community are, for now, two different things.

Taken together, the release fits a pattern forming around decision models: rather than one company's classifier becoming a default, multiple well-resourced players — an infrastructure startup, a CDN operator, an open-weight model lab, and now a search company — are each shipping their own typed-decision model under permissive licenses (pplx-decider ships Apache-2.0), competing less on branding than on benchmark placement within community-run leaderboards like JevBench. Whether that produces a genuinely standardized interface for agent routing, moderation, and rubric-based grading, or simply a fragmented field of similarly-shaped models each claiming the top spot on a different slice of a shared benchmark, is still an open question the download numbers have not yet answered.