pplx-decider on Ollama: Perplexity's Silent 27B Decision Model Becomes Runnable

Perplexity published the weights of pplx-decider-v1.1-27b on 2026-10-05, and the first checkpoint it superseded, pplx-decider (repo name pplx-decider-v1-27b), on 2026-10-01. Both are fine-tunes of Qwen3.8-27B that answer in probabilities instead of sentences, and both sit in the same System One lineage as TypeSafe's Jev 1.13. What is new is not the model: as of today, pplx-decider is an entry in Ollama's own model library, installable with one command. Before that listing appeared, running this checkpoint meant a CUDA GPU with room for roughly 49 GiB of bf16 weights. Now it is a 17 GB quantised pull on a laptop. That single change is what this page is about, and the model it makes runnable is fourteen days old.

What changed, in concrete terms

The Ollama library entry is named pplx-decider, and the library has been steadily collecting decision models — Laya, Nimble, Kev, Clef, Strands Decider — through the autumn. This one is the newest addition, and on the maintained registry it exposes three tags (pplx-decider:latest, pplx-decider:27b and pplx-decider:27b-q4_K_M) that resolve to the same manifest digest. A single artifact, not a size ladder.

• Format — GGUF served by llama.cpp, versus the bf16 safetensors shards on Hugging Face

• Download size — about 17 GB at q4_K_M, versus roughly 49 GiB of weights plus working memory for the bf16 checkpoint

• Vision layer — the manifest carries a separate ~927 MB projector alongside the model layer, so image input is packaged with the text weights

• Hardware — a consumer machine rather than a CUDA box

The interface is Ollama's /v1/systemone endpoint, which it added in v0.35.0 and extended for Clef models in v0.35.1. It has also just grown a command-line front end: v0.40.4, released 2026-10-12, added decision models to ollama run, so a yes/no question, a label choice or a rubric score can be asked straight from a terminal with --ask, --choice and --level. That release is three days old; the model it makes convenient is two weeks old.

What the repository actually documents

The checkpoint is not a chat model with a terse prompt. Its card describes a native decision layout: the Qwen3.8-27B backbone at the released class name, plus a separate readout.safetensors holding a bf16 [255, 5120] decision head — a 255-way readout over the option slots, not a full-vocabulary language head. There is no decoding loop. Every option of every question is scored in one forward pass, which is why the response reports zero output tokens.

Three details in the card matter more than the benchmark table, because they are the ones that break a naive integration:

• Attention mode. Full-attention layers use a saved noncausal mode, and the card says plainly that default causal inference does not reproduce the evaluated setup. Launch the checkpoint with an off-the-shelf causal loader and you get numbers that are not the published ones.

• Calibration temperature. The supplied DecisionModel.predict applies a saved temperature. If you read logits directly, apply it exactly once — applying it twice, or not at all, mis-scales every probability you then branch on.

• A real decision limit. Text and all images for one decision must fit inside an 8,192-token budget, and over-limit inputs are rejected rather than truncated. Training and the reported evaluation capped images at 512×512; raising the cap is now configurable via max_image_pixels, defaulting to 2048×2048, but the card warns that higher resolutions sit outside the training distribution.

The repository is otherwise unusually complete for a model nobody announced: eleven safetensors shards, a decision_config.json, a chat template, and a release-manifest.json recording file checksums and checkpoint provenance. Licence is Apache 2.0.

Two pieces of the documentation have drifted. The v1.1 card still tells you to authenticate against "this private repository" — the repository reads publicly today. And the older v1 card carries a new_version pointer at v1.1, which is the honest signal that v1 is superseded.

The one thing to check before you plan around it

The Ollama listing advertises a 256K context window. That number appears to be inherited from the base model — Qwen3.8-27B does carry 262,144 positions — and it is not the contract the decision checkpoint documents for itself. The checkpoint rejects any decision whose text plus images exceed 8,192 tokens. Plan a pipeline on 256K, feed it a 40,000-token contract plus four page images, and the request fails rather than getting cut down. The larger figure describes the architecture; the smaller one describes the API you are actually calling.

Image support has a similar edge. The Ollama model card for pplx-decider shows an image request against /v1/systemone, but Ollama's own capability page for that endpoint states images are "only for Clef models." One of those two documents is stale. Treat multimodal input on this path as unverified until you have run it yourself — the underlying checkpoint clearly takes images, and the manifest clearly ships a projector, but the runtime's documented gate is narrower than the model's card implies.

A vendor number, and a different number

Perplexity's own evaluation for the v1.1 checkpoint reports a Decision Index of 61.56 against 57.9 for Jev on the same suite and 56.4 for the checkpoint it replaced, with the biggest category gains in language (69.45) and retrieval (61.26). Those figures are vendor-reported and were produced on Perplexity's side; for the earlier checkpoint the card goes further and says the results were measured through the Perplexity API. That is a legitimate thing to publish and a weak thing to rely on.

The independent picture is more useful. An open community leaderboard for this model class, the Jev Decision Index, has been scoring pplx-decider-v1.1-27b on its own panel and its 2026-10-14 data release reports the checkpoint completing 110,179 of 110,201 scored cases across 42 benchmarks on a single RTX PRO 6000. On that board the model lands a balanced skill score of 62.75 and a rank of sixth, with Jev twelfth at 60.11. Two things the vendor table does not show fall out of the same data:

• Where it is strong. ARC-Easy 99.3, ARC-Challenge 97.0, function-call yes/no decisions (BFCL) 97.2, FinEntity 95.4, BPoMP 95.2, CLINC150 91.5. Classic classification and retrieval tasks are where this checkpoint earns its keep.

• Where it is not. Humanity's Last Exam 16.2, ChessBench 17.7, SATA-Bench 28.2, ACOS 28.2, the schema-guided dialogue set 33.2, BRIGHT 50.9, GPQA Diamond 56.1. Hard multi-step reasoning is not what a one-pass readout buys you, and a stack that quietly routes reasoning work to a decision model will produce confident wrong branches.

Calibration deserves its own line, because it is the whole point of a model that returns probabilities instead of labels. On 34,829 independently scored decisions the checkpoint averaged 76.1 percent accuracy against 82.9 percent mean confidence — an expected calibration error of 0.071 and a Brier score of 0.335, with 2.0 percent of cases sitting above 0.95 confidence and still wrong. That is a usable but not flawless probability surface, and it is measurably better calibrated than Jev's 0.074 ECE on the same board. If you are setting a 0.9 threshold and treating it as near-certainty, that 2 percent is the number to design around.

Latency on the same hardware came in at a 104 ms median and an 835 ms p95 across 110,000 decisions, which is the practical argument for a decision model over a small language model in a gating position: the p95 is the budget, not the mean.

What a decision costs

Perplexity runs its own hosted Decisions API for these checkpoints — POST https://api.perplexity.ai/v1/decisions, with pplx-decider-v1.1-27b as the default model ID and the earlier checkpoint still addressable. That pricing page lists $0.02 per million input tokens, output tokens free, and no per-request fee, with a ceiling of 128 questions per request, 262,144 input tokens and 10 requests per second per organisation.

Against that, the local path costs nothing per decision once the 17 GB have landed. For a triage queue running a handful of typed questions per ticket, that arithmetic decides itself — and it is worth noticing that the hosted ceiling of 262,144 input tokens is not the limit your local copy will enforce, because the checkpoint's own budget is 8,192.

Where the local route is not the answer is the rest of the stack around the decision. This is a 27B model that returns one probability per option; it will not draft the reply, summarise the ticket or call the tool. If you want the decision and the writing behind a single key, that is what OrcaRouter is for — one API across 200+ models with provider list price passed through at no per-token markup, plus the routing DSL when you want a decision model and a writing model composed into one call. Two pieces of this particular story are already on the platform: the base checkpoint is served as Qwen3.8-27B, so you can compare the fine-tune against its own foundation on the same key, and the reference model it is measured against is served as Jev 1.13 through the same systemone endpoint shape. Perplexity's decider is not one we carry — it is available through Perplexity's own API and several third-party platforms — so if you need it specifically, the local pull above and Perplexity's endpoint are your routes.

Three questions the repository does not answer

Which checkpoint is in the quantised artifact? The Ollama library entry is named for the family, not a version, and its benchmark block links the v1 repository — the 11-benchmark table reads 85.7 percent, which is the v1 card's overall figure and not v1.1's Decision Index suite. v1.1 is the stronger checkpoint by Perplexity's own accounting, and it is not obviously the one you are pulling. Check the digest before you build on the assumption.

Did Perplexity intend any of this to be public? No announcement, blog post, changelog entry or API-availability notice for these weights has surfaced. The checkpoints have been read by thousands — 2,740 downloads on v1 and 3,154 on v1.1 at the time of writing — but attention on a model card is not the same as a vendor saying it is supported. Treat the public repository as the authoritative artifact and everything else as inference.

Is a vendor-hosted, officially announced version coming? Unknown. The existence of a documented, priced Decisions API on Perplexity's side suggests the hosted route is the intended product and the weights are upstream of it. But that is a guess about intent, not a fact from a document, and it is labelled as one.

Who should move, and who should not

If your pipeline already branches on typed provider decisions — ticket routing, refund detection, agent step gating — this is a two-week-old checkpoint with independently reproduced scores, an Apache 2.0 licence and, as of today, a 17 GB path onto hardware you own. The evaluation suggests it is better calibrated than the reference model and roughly two ranks better on a public board, and the local copy removes the per-decision price entirely. Pull it, run it against your own labelled examples, and compare it to whatever you use now. The 8,192-token decision budget is the constraint to test against first, not the 256K the listing advertises.

If you need reasoning rather than classification, wait. The independent panel puts this checkpoint in the low teens on hard knowledge benchmarks, and a probability surface that is confident on a question it cannot actually answer is worse in production than a slow model that thinks first. The case for a decision model is the 104 ms median and the fact that sixty questions cost one pass — and none of that helps where the answers are wrong.

Running the decision step on your own hardware is also a good time to look at what has already gone wrong in this class: the documented AI incidents and agent failures involving decision models are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.