AI ExplainersNov 9, 20269 min read

What Is pplx-decider? Perplexity's Open Decision Model, Tested

pplx-decider is the most accurate decision model you can run yourself: 92.2% on DecideBench v2, 2.2 points behind Drex Max, at four times its cost on our GPU.

On this page
  1. What is pplx-decider, and who makes it?
  2. Which versions of pplx-decider exist?
  3. How did pplx-decider score on DecideBench?
  4. What does pplx-decider cost to run yourself?
  5. pplx-decider vs Jev and Drex Max: which is better?
  6. How do you run pplx-decider yourself?
  7. What we did not test
  8. FAQ
  9. Sources

pplx-decider is Perplexity's System One model: it reads an input, a question and options, and returns a probability for each option. The weights are open, and the same model powers Perplexity's hosted Decisions API. On 9 November 2026, our DecideBench v2 test ran pplx-decider v1.2 27B on our hardware. It scored 92.2%, the best open model, 2.2 points behind Drex Max (94.4%), the best hosted one.

  • Who makes it: Perplexity, the AI search company. The weights are free to download under the Apache 2.0 licence.
  • Sizes: a 27B model (pplx-decider-v1.2) and a 4B model (pplx-decider-flash-v1.2).
  • Our score: the 27B model got 92 of every 100 decisions right. Flash got 78.
  • Our cost, self-hosted: $386.34 per million decisions on one rented NVIDIA graphics card.
  • Perplexity's API price: $0.02 per million input tokens for the 27B model. We did not test the API.

If "System One model" is new to you, what is a System One model explains the category. For every model we ranked, see the best System One models of 2026.

What is pplx-decider, and who makes it?

pplx-decider is a family of decision models from Perplexity. You send it a state (text, JSON or images) and named questions, and it answers each one with probabilities instead of written text.

Perplexity's Decisions API quickstart lists three kinds of question:

Question typeWhat you askWhat comes back
noulYes or noThe probability of yes
choiceOne of your optionsA probability per option, the top option, a confidence
scoreA level on your rubricA probability per level and an expected score

DecideBench tests the choice kind: pick one option, such as approve this refund or send it to a person. Perplexity's API accepts up to 255 options per question.

The weights live on Hugging Face, the model-hosting site, under perplexity-ai. Each model starts from an open Qwen language model. Perplexity swaps the part that writes text for a small part that scores options. The model never writes a reply, so the hosted API bills only for the text you send in.

Which versions of pplx-decider exist?

Perplexity has shipped three versions of the large model and one small Flash model. Perplexity's API changelog dates all three API releases to October 2026, without days. The model cards give no release dates either, so we don't give one.

NameSizeWhat the model card says changedPerplexity's API today
pplx-decider-v1-27b27BFirst release, fine-tuned from Qwen3.8-27BRouted to v1.2
pplx-decider-v1.1-27b27BLifts the causal mask, adds training data including tasksourceRouted to v1.2
pplx-decider-v1.227BJoint request format, reworked training data, several training runs averaged ("model souping"), images up to 1.6 MP$0.02 per million input tokens
pplx-decider-flash-v1.24BDistilled from the 27B models, built on Qwen3.5-4B$0.01 per million input tokens

pplx-decider-v1-27b opened the Decisions API at $0.04 per million input tokens. Its model card claims 85.71% overall against Jev's 84.51% across 11 public benchmarks, "measured through the Perplexity API". That is Perplexity's claim, not ours. Jev led on 6 of the 11 rows. The card does not say how it weighted the overall figure.

pplx-decider v1.1 replaced v1 in the API and cut the price to $0.02. Its card reports a Decision Index score of 61.56 against 56.4 for v1 and 57.9 for Jev, again Perplexity's own numbers.

pplx-decider v1.2 is the current 27B model. Its card reports a Decision Index score of 68.09, up from 62.30 for v1.1 on the same suite. The card also notes that some training data overlaps the Decision Index benchmarks and was decontaminated against the test sets.

pplx-decider-flash v1.2 is about a seventh of the 27B model's size. Its card recommends it "when latency or cost matters more than the last points of accuracy" and reports 58.88 on Decision Index 0.3. Hugging Face's file viewer lists 5B parameters; Perplexity calls it 4B.

Old names still work on the API. Requests for pplx-decider-v1-27b or pplx-decider-v1.1-27b are routed to v1.2 and billed at $0.02, and pplx-decider-latest follows the newest 27B version.

One naming trap: Decider-2B and Decider-4B on our leaderboard are not Perplexity's. They come from a Hugging Face user called Mapika, whose card never mentions Perplexity. They scored 31.8% and 63.6%.

How did pplx-decider score on DecideBench?

The 27B model got 92 of every 100 decisions right. That put it third of the 26 decision models, level with the second-best hosted API. Flash got 78 of every 100, tied for eleventh.

DecideBench v2 is our benchmark: 450 multiple-choice decisions in nine jobs. Each input looks like what the model would see at work, such as a raw email with its headers. Every decision has a twin that differs by one small edit that flips the right answer. "Pair accuracy" counts the twins where the model got both right, so a model that answers from topic words alone loses points there. Our benchmark post covers the method.

ModelRight answersPairs both rightCost per millionMedian wait
Claude Opus 4.8 (reference LLM)99.3%98.7%$26,542.701,336 ms
Drex Max (Nace.AI, API)94.4%89.3%$91.00258 ms
REAX 1 27B (Decision Models, API)92.4%87.1%$120.29545 ms
pplx-decider v1.2 27B (self-hosted)92.2%86.2%$386.341,528 ms, no network
s1-pro (Decision Models, API)91.6%85.3%$120.93418 ms
Jev (TypeSafe, API)86.2%72.9%$75.04334 ms
yev1-4b (XY Space, self-hosted)79.3%62.7%$71.15275 ms, no network
pplx-decider-flash v1.2 4B (self-hosted)78.2%62.2%$125.34284 ms, no network

Claude Opus 4.8 is the reference the test was built against, not a decision model. yev1-4b is ours; it sits in the table because it is the nearest small model to Flash.

On 450 decisions, one point is about 4.5 answers. REAX 1, pplx-decider and s1-pro sit within 0.8 points of each other, so read them as a cluster.

The job-by-job scores show where the 27B model loses its points:

JobOpus 4.8Drex Maxpplx-decider 27BJevpplx-decider Flash
Approving an AI agent's proposed action100%100%98%94%84%
Routing an agent to a tool98%96%98%94%92%
Checking a claim against its sources98%98%98%94%90%
Moderating a post100%100%100%96%94%
Triaging an inbound email98%98%96%88%74%
Ruling on a return100%90%86%76%68%
Gating a staged software release100%74%64%60%40%
Reading a customer's intent100%100%100%96%96%
Triaging a bug report100%94%90%78%66%

The release gate is the weak spot. Those decisions need a crash-free rate or an error ratio computed from metrics and checked against an ordered list of rules. pplx-decider got 64% of them right, against Opus 4.8's 100%. No decision model passed 74%. If your decisions depend on arithmetic over a written policy, test that job before you rely on any of them.

What does pplx-decider cost to run yourself?

On one rented graphics card, pplx-decider v1.2 27B cost $386.34 per million decisions. That is about 5.1 times Jev's API price and 4.2 times Drex Max's, for a model that ranks between them on accuracy.

We price self-hosted models by the time they hold the graphics card. All 450 decisions took 186.1 seconds on one NVIDIA RTX PRO 6000 Blackwell card. We rented it as an AWS g7e.2xlarge at $3.36 an hour, with 4 requests in flight. Flash finished the same 450 in 60.4 seconds. A server kept busier than ours would cost less per decision.

Self-hosted wait times were measured on the GPU box, so they include no network; hosted ones do.

Perplexity ships pplx-decider as a Python library, not a server. So we wrote a small server, `tools/pplx/serve.py`, to call it like the other models. Requests that arrive while the model is busy wait in a queue and are scored together in one call. Batching mattered:

pplx-decider v1.2 27BRight answersCost per millionMedian wait
Unbatched92.4%$8753,443 ms
Batched (published score)92.2%$386.341,528 ms

Batching cut the bill by more than half for 0.2 points of accuracy, about one answer in 450.

Perplexity's hosted price is a different kind of number. Its pricing page charges $0.02 per million input tokens for v1.2 and $0.01 for Flash. Output is free.

We did not call Perplexity's API, so we have no hosted accuracy, speed or cost of our own. Perplexity's example request has 600 input tokens. Ours carry worked examples and run longer, so the two prices don't line up. System One model pricing compares the hosted price lists.

pplx-decider vs Jev and Drex Max: which is better?

On our test, pplx-decider 27B beat Jev by 6 points and trailed Drex Max by 2.2. The pair scores widen the gap over Jev to 13.3 points.

pplx-decider v1.2 27BJevDrex Max
Right answers92.2%86.2%94.4%
Pairs both right86.2%72.9%89.3%
Cost per million (as we ran it)$386.34, GPU time$75.04, API$91.00, API
Median wait1,528 ms, no network334 ms258 ms
Weights you can downloadYes, Apache 2.0NoNo
pplx-decider against Jev and Drex Max on 450 decisions
pplx-decider against Jev and Drex Max on 450 decisions
Right answersBoth halves of a pair right
Jev86.2%72.9%
pplx-decider v1.2 27B (self-hosted)92.2%86.2%
Drex Max94.4%89.3%

Source: DecideBench v2, measured 9 November 2026

If your data has to stay on your own hardware, pick pplx-decider 27B. It is the only self-hosted model we tested that beat Jev.

If you only need the most accurate answer per dollar, pick Drex Max. It is more accurate, faster and cheaper than pplx-decider as we ran it.

If you are moving off Jev and want open weights for later, start with pplx-decider. Perplexity's card claimed a narrow lead over Jev. On our 450 decisions the lead was wider.

We called Jev through AI Space, XY Space's platform, for these runs. For background, see what is Jev. OpenAI sells a rival under the same API name; what is OpenAI's Decisions API covers it. Its model, GPT-6 Luna, scored 84.0%.

How do you run pplx-decider yourself?

Download the weights from Hugging Face, install the bundled Python package, and call decide() on an NVIDIA graphics card. The v1.2 model card lists what you need:

  • Python 3.12 or later.
  • About 52 GiB of free graphics-card memory for the 27B weights, plus working room. Flash needs about 9 GiB.
  • A Hugging Face access token, then snapshot_download for the repository.
  • pip install ./source from the snapshot, or requirements.txt. The faster path adds requirements-fast.txt, which needs flash-linear-attention and torch 2.10 or later.
  • Load with Decider.from_pretrained(<snapshot>, device="cuda"), then pass decide() a state and a dictionary of questions.

Very long inputs need more memory. The card reports a peak of about 107 GiB for an input of about 243,000 tokens. Our inputs are far shorter, and both models fit on our one card.

To serve it over HTTP the way we did, our wrapper exposes the model behind Jev's /v1/systemone route and batches up to 16 queued requests per call:

python3 tools/pplx/serve.py --model <pplx-decider snapshot> --name pplx-decider-v1.2 --port 8790

We ran the fast path (fast=True) with Perplexity's own package at the pinned snapshot recorded in our results files. Open source System One models compares pplx-decider with the other models you can run yourself.

What we did not test

We tested two models, v1.2 27B and Flash v1.2, self-hosted, with one worked example per option. We did not test v1 or v1.1, Perplexity's hosted API, its web-grounding mode, image inputs, or the noul and score question types. The test's items, examples and policies were written by a language model and checked by three Claude models, and 450 items leave wide margins on any single job.

FAQ

Is pplx-decider open source?

The weights are. Every pplx-decider model card on Hugging Face lists the Apache 2.0 licence, and each snapshot bundles the Python source used to run it. The v1.1 card says some of the added training data comes from the tasksource project. The card also asks for a Hugging Face access token to download.

What is the difference between pplx-decider and pplx-decider-flash?

Size, speed and accuracy. pplx-decider v1.2 has 27B parameters. Flash v1.2 has 4B and was trained to copy the larger models. On DecideBench v2 the 27B model got 92.2% right, and Flash got 78.2%. On the same graphics card, Flash cost about a third as much and answered about five times faster. On Perplexity's API, Flash costs half as much per input token.

Is pplx-decider better than Jev?

On our test, yes. pplx-decider v1.2 27B got 92.2% of 450 decisions right, against Jev's 86.2%. On pairs it led by 13.3 points. Run on one rented graphics card it cost $386.34 per million decisions, about five times Jev's $75.04 API price, and answered slower. Flash, at 78.2%, scored below Jev.

What GPU do you need for pplx-decider?

Perplexity's card asks for an NVIDIA graphics card with about 52 GiB of memory free for the 27B model's weights, plus working room. Flash needs about 9 GiB. We ran both on one NVIDIA RTX PRO 6000 Blackwell (AWS g7e.2xlarge, $3.36 an hour on demand). Very long inputs need more: the card reports about 107 GiB at 243,000 tokens.

Is pplx-decider-v1-27b still available?

On Hugging Face, yes: the v1 card still hosts the weights and points to v1.1. On Perplexity's API, requests that name pplx-decider-v1-27b or pplx-decider-v1.1-27b are routed to v1.2 and billed at $0.02 per million input tokens, according to Perplexity's pricing page.

Sources

Written by

Cho Yin Yong

Principal AI Solutions Engineer, XY Space

Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.

More from Cho Yin Yong

Share this article

Work with us

We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.

Start a project
Work with us

Book a call.We'll come back with specifics.

Start with the map of your organization, or with the one job that hurts. Measured in hours and money, and everything we build stays yours.

Loading form…