pplx-decider-flash on Ollama: Perplexity's 4B decision model, first look

Perplexity's 4B decision model is on the Ollama library as of today, with its own page, a 5.6GB download and a 256K context window. Here's what it is, what hardware it wants, and how to call it.

Listen to this post · 9 min

Accessibility: the whole post is read aloud — Space plays or pauses, ← → skip 10 seconds, and you can change the speed. It keeps playing with your screen locked.

Share this post

In short

  • pplx-decider-flash is Perplexity's 4B decision model, distilled from the full-size pplx-decider — same request format and image support at about one seventh the size, with a 5.6GB download and 256K context.

Perplexity's pplx-decider-flash is on the Ollama library as of today (3 November 2026), with its own page at last: a 4B decision model built on Qwen3.5-4B and distilled from the full-size pplx-decider. It doesn't generate text. You give it a state — text, JSON, or images — plus up to 64 typed questions, and it returns a calibrated probability for every answer in a single forward pass, with zero output tokens. The download is 5.6GB and the licence is Apache 2.0.

If you read our October piece on the parent model, this is the 4B sibling we flagged when its size wasn't yet stated on the page. It now has one.

What pplx-decider-flash is

The pitch is simple: the same request format and the same text and image support as pplx-decider, at about one seventh the size. Where the parent model is built for bigger machines, flash is sized for the hardware most people already own.

The mechanics work like the other decision models we've covered this autumn — kev, d1, Strands Decider and the rest. Instead of writing prose, the model scores options. A support ticket goes in; "is this urgent?" comes back as a probability, along with "which team should handle this?" as a probability distribution over your named options, and "how severe is it?" as a probability-weighted rubric level. One forward pass, no output tokens, up to 64 questions per request.

That last point is the latency story. Because nothing is generated, there's no autoregressive decoding loop — the entire cost is the prompt pass. Perplexity positions flash for cases where latency or cost matters more than the last points of accuracy, which is a fair description of most routing and triage workloads.

Sizes, memory and the hardware you need

There's one size on Ollama, and the tags are identical packages:

TagDownloadContextInputs
pplx-decider-flash:latest5.6GB256KText, Image
pplx-decider-flash:4b5.6GB256KText, Image

The page doesn't state memory requirements, so here's the reasoning rather than a spec: budget roughly the download size in RAM or VRAM, plus KV cache for your context. For short states and a handful of questions, 8GB of VRAM or unified memory is a realistic floor; if you actually fill the 256K window — long documents, big batches of images — the cache grows and you'll want meaningfully more. A decent laptop with 16GB can hold it for typical use; a gaming GPU with 8GB or more is fine for the workloads this model is built for.

The 256K window is the other number worth pausing on. It's unusually large for a small decision model. Strands Decider, the 2B we covered in October, sits at 4,096 tokens — fine for a short ticket, useless for routing on long documents. Flash can take a whole contract, a long thread, or a batch of screenshots and still have room for your questions.

How to run it locally

Pull it first:

ollama pull pplx-decider-flash

Then the CLI. The pattern is --ask plus the question, with the state on stdin or following the model name. There are three question types:

# Yes/no (noul)
ollama run pplx-decider-flash --ask "Is a refund being requested?" < ticket.txt

# Choice between named options
ollama run pplx-decider-flash --ask "Which team should handle this?" \
  --choice billing="Payments and refunds" \
  --choice technical="Outages and errors" < ticket.txt

# Rubric levels, lowest first
ollama run pplx-decider-flash --ask "How urgent is this?" \
  --level "Routine" \
  --level "Soon" \
  --level "Immediate" < ticket.txt

For anything programmatic, use Ollama's /v1/systemone endpoint — the same one the other decision models use. Put the text to judge in state (a string, or a JSON object or array for structured input), optionally base64-encoded PNG, JPEG or WebP images shared by all questions, and 1 to 64 named questions. Each question has a type (choice, noul, or score), instructions, and usually criteria describing the options or levels. Ollama builds the prompt for you.

curl http://localhost:11434/v1/systemone -d '{
  "model": "pplx-decider-flash",
  "state": "Our app crashes on launch for every user since this morning'\''s update. Please fix it now.",
  "questions": {
    "urgent": {"type": "noul", "instructions": "Is this support request urgent?"},
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this request?",
      "criteria": {
        "billing": "Payments, invoices, and refunds",
        "technical": "Outages, errors, and configuration"
      }
    }
  }
}'

Answers come back in the same order as the questions, each with probabilities and a confidence value from 0 to 1. One thing the readme is careful about: confidence shows how strongly the probabilities favour one answer, not the chance the answer is right. A model can be confidently wrong. If you're gating agent actions on these numbers, calibrate against your own data before trusting them.

There's also a Python route via TypeSafe's SDK pointed at Ollama. As of today, the Ollama page is the only run documentation we've seen — no llama.cpp, LM Studio or vLLM instructions.

What it's good for — and what it can't do

The readme spells out four patterns:

  • Route a request. Define the destinations and when each applies; get back the chosen destination and each destination's probability.
  • Check a condition. A yes/no question; get true or false with the probability of each.
  • Grade an answer. A rubric of levels; get the probability-weighted level and each level's probability.
  • Gate an agent step. A condition plus the agent's proposed action; get go or no-go with probabilities.

Because it's multimodal, the state can be screenshots, receipts and forms, not just text — so "does this receipt show a refund-eligible item?" or "which field is missing on this form?" are in scope.

The limits are the flip side of the design. It never writes prose, so it's not a chat model, a summariser or a drafting model. If your pipeline needs text out — a reply to the customer, a summary of the ticket — pair it with a small chat model and let flash make the structural decisions. Its Qwen3.5-4B base is a decent pointer to the class of chat model that pairs well. That division of labour is usually cheaper and faster than asking one model to do both.

And the usual caveat applies: all quality numbers are Perplexity's own evaluations. We've seen no independent benchmarks yet.

How it compares: flash vs pplx-decider vs the rest

Perplexity's evaluation puts flash at 58.88 overall on the Decision Index 0.3 (60.05 with one question per request) and 72.53 on the vision suite, against 68.09 and 77.43 for the full-size pplx-decider. So you give up roughly nine points on text and five on vision, and get a model about a seventh the size. Whether that trade is right depends on your volume: at high request rates, the difference in cost and latency compounds fast.

Here's how the field looks as of today. Every figure in the rows below comes from our earlier coverage of each model, linked in the table — follow those for the full details and the caveats:

ModelMakerSizeDownloadContextVisionNotes
pplx-decider-flashPerplexity4B5.6GB256KYesDistilled from pplx-decider
pplx-deciderPerplexity27B17GB256KYesTop scores in the family
d1Liquid AI3B3.7GB32KNoText-only, smaller window
kevJared Palmer0.8B–27B812MB–18GB—NoMac-only (MLX runner)
d3vLLM Semantic Router2b–27b3.62–28.61GB256KYesVision at every size
Strands DeciderStrands2B1.3GB (NVFP4)4KNoShort documents only

Within the Perplexity family the choice is straightforward: if you have the VRAM for the parent model and want the top scores, run pplx-decider. If you're on a laptop, a small GPU, or serving many requests, flash gives you the same API and format at a fraction of the cost.

Against the wider field: flash is the only 4B decision model we've covered with both vision and a 256K window. d3 has both too, but starts at 2b with lower publisher-reported scores; kev is Mac-only; d1 has no vision and a 32K window; Strands Decider's 4K context rules out long documents. One honest caveat: these are publisher-reported numbers on different internal benchmarks, not a like-for-like comparison. Treat the table as a shortlist, not a ranking.

Worth trying if you're building a router or gate

If you want to experiment with decision-model patterns from code before committing a machine, the marketplace has small open models callable through one OpenAI-compatible API — useful for prototyping the routing loop, then moving it local. And if you have a GPU or Mac that holds 5.6GB comfortably and would otherwise sit idle, hosting open models can earn from each request.

Questions people ask

What is pplx-decider-flash?

It's Perplexity's 4B decision model, built on Qwen3.5-4B and distilled from the full-size pplx-decider. You give it a state and up to 64 typed questions, and it returns a calibrated probability for every answer in a single forward pass — it never generates text.

Models in this post — live on Inpherio

Have a GPU, a Mac or a spare PC?

Let it earn while you sleep.

Run open models for people who need them. You set the price and keep everything but the 15% fee.

Become a host

Building with AI?

Every open model, one API.

OpenAI-compatible. Priced by the people who run the hardware.

Start free See the marketplace

Read “pplx-decider-flash on Ollama: Perplexity's 4B decision model, first look” at https://inpherio.co.uk/blog/pplx-decider-flash-on-ollama-perplexity-s-4b-decision-model-first-look