Perplexity Decider: free decision model runs locally on your Mac

What's new

Perplexity has launched Decider for Mac. According to independent coverage, it was announced on October 22, 2026, and Perplexity's own forum post is dated October 23. It is a small app in the Mac menu bar that keeps the pplx-decider-flash-v1.2 model loaded in memory and runs it locally. No API key (a kind of password for the service) is required.

The model is not a chatbot. It has 4 billion parameters (roughly speaking, the model's internal tuning knobs) and can answer questions about text or images. Instead of a long written reply, it responds with probabilities. There are three question types: yes/no, choosing between several options, and scoring on a scale.

The app requires a Mac with Apple silicon (M1 or newer), macOS 13 or later, 16 GB of memory and about 6 GB of free disk space. Windows, Linux and older Intel-based Macs are not supported. The app is downloaded as Decider.dmg directly from Perplexity. The 5.8 GB model is downloaded on first launch, and a Hugging Face token may be required.

What's smart

Your content stays on the Mac. That means you can have the model sort files that you are not allowed or do not want to send to an online service, such as internal documents or images.

The app also opens a local API at 127.0.0.1:8080/v1/decisions. It is a door that other programs on the same Mac can use. Requests use the same format as in Perplexity's hosted service, just with a different address and without an authorization header. That makes it easy for a developer to move a task from the web to the Mac. Perplexity recommends one question per request. You can set the port and have the app start at login.

Perplexity says the Mac version is a quantized version (q8_0 GGUF, in other words a compressed edition) and is not quite the same as the hosted model.

Cheaper or better?

The app is free, with no pay-per-use charges. By comparison, Perplexity's hosted Decisions API costs USD 0.01 per million input tokens (chunks of text) for the small model and USD 0.02 for the large model with 27 billion parameters. Output is free, and the context window is 256,000 tokens. The price is already low, so the main benefits of the local version are privacy and control.

Benchmark figures come from an independent report that has not been checked against Perplexity's own sources, so they should be treated with caution. According to that report, the model scores 58.88 on the Decision Index with grouped questions and 60.05 with single questions. On images it scores 72.53, compared with 77.43 for the 27B version. Calibration accuracy is 88.8 percent across 1,497 test questions. The report also mentions competitors such as TypeSafe's Jev, NVIDIA/Stanford's CLM-8B, Cloudflare's Clef, AWS Strands Decider 2B and Microsoft Decision 1, but without a direct comparison here. Perplexity has previously said itself that the large model is the highest-scoring on ClaimBench. That is Perplexity's own claim.

The license for the app and the model weights is not specified. Free therefore does not necessarily mean open. The sources also do not say whether there are regional restrictions in Denmark or the EU, but none have been found.

What it's good for

Decider suits tasks where you have many items to sort and want a quick, short answer: Is this image relevant? Which of three categories does the document fit into? How serious is this inquiry on a scale? Because everything runs locally, it is especially interesting for companies and employees with sensitive files.

It is not for conversation, writing or anything else where you want free text. It also requires a relatively new Mac with plenty of memory, so many older or cheaper computers are ruled out.