Perplexity Hybrid Compute pairs a cloud orchestrator model with a local model on your Mac, and a privacy gate keeps sensitive files on your device. It shipped in October 2026, and a full setup takes about 10 minutes. This walkthrough covers setup, the privacy gate, the model pairings, and a custom skill that moves most work to the local model to save cloud credits.
What Is Perplexity Hybrid Compute?
Perplexity Hybrid Compute is a Mac app feature from Perplexity, the AI search company, that combines a cloud orchestrator model with a local model running on your own device. It launched in October 2026 and ships with a pre-installed privacy gate that scans conversation input for personal information. When it finds something sensitive, the app flags it in the chat and asks how you want to proceed.
The speaker in the source video, testing on a 24 GB M6 Mac mini, describes it as the fix for his biggest hesitation about giving AI agents access to personal files such as client invoices, private meeting notes, or a personal diary. Everything lives in a single app: you download the Perplexity desktop app, download the local model inside it, and enable privacy mode. He reports the full process takes about 10 minutes.
The architecture matters for understanding the rest of this article. Every new conversation starts with a cloud orchestrator, such as GLM 5.3 in the video. That orchestrator can delegate work to a local sub-agent on the Mac. The orchestrator is fixed for the conversation, but the local model stays available throughout, loading on demand only when a task actually needs it.
How to Set Up Hybrid Compute on a Mac
Setup runs through the Perplexity desktop app in five short steps, and takes about 10 minutes end to end. A free Perplexity account works for the app itself, but the local-model handoff requires a paid plan, reported at $20 per month in the video. Here is the sequence the walkthrough follows:
- Download the Mac app from the Perplexity Hybrid Compute page and run the onboarding assistant.
- Create or sign in to a Perplexity account; a free account starts the app, a paid plan unlocks the local handoff.
- In settings, enable "sensitive data on your Mac" to install the privacy gate, a separate small model that runs entirely on your machine.
- Open the orchestrator settings, choose configure, and download a local model to the device.
- Pick your default cloud orchestrator model and turn Hybrid mode on.
The app also has a mobile companion, so you can drive the agent on your Mac mini or Mac Studio from your phone. The cloud orchestrator runs remotely either way; the Mac is where the local sub-agent lives.
How the Privacy Gate Decides What Stays Local
The privacy gate is a separate, small model trained to recognize personally identifying information: names, email addresses, phone numbers, account numbers, and similar data. It runs entirely on your Mac and inspects files and messages during a conversation.
When it detects sensitive content, a pop-up appears in the chat with two choices: route that part of the conversation to the local model on your Mac, or send it to the Claude Code if you are comfortable sharing. The default behavior for an uploaded folder is to pass every file through the gate before anything reaches the cloud.
The video demonstrates this with a folder of invoices covering a fertility clinic, tax consultancy, family law, and payroll, the kind of files that should trip the detector every time. Dragging one psychology invoice into the chat triggered an immediate local flag. The app listed exactly what the small model had extracted, and after choosing "process on my Mac," the summary came back with a note that the file contained personal information and was kept on the computer.
Batch behavior has a friction point worth knowing before you rely on it. When the speaker uploaded a folder of six sensitive PDFs at once, the gate flagged the first invoice and then presented a pop-up for each subsequent file, though you can select "don't ask again this session." Activity Monitor showed the local model loading under demand inside a Python process, using the full GPU and about 66% of system memory during the batch run.
Which Local and Cloud Models Can You Pair?
Hybrid Compute lets you choose both halves of the pair. On the Claude Code, the app lists multiple orchestrators, and newly released models appear as they become available. On the local side, the model downloads directly inside the app.
The speaker's cloud list included GPT 6.1 Soul, which he notes had just been released around OpenAI Dev Day, plus ChatGPT models, Claude models, GLM 5.3, Kimmy K3, Grok, and Deep Seek V4 Pro. He settled on GLM 5.3 as his default because it consumes fewer credits and handles complex reasoning well. One clarification for readers comparing tools: Claude models in this list are Anthropic chat models, which is a different product from Claude Code, Anthropic agentic coding tool that runs in the terminal.
The following comparison reflects the configuration shown in the October 2026 walkthrough on a 24 GB Mac mini. Model listings and hardware requirements come from the speaker's on-screen app, so treat them as observed in that setup rather than as a general product spec.
| Component | Model shown | Role | Notes from the video |
|---|---|---|---|
| Cloud orchestrator | GLM 5.3 | Plans tasks, delegates work | Default in the video; lower credit use |
| Local model | Gemma 4 E4B | Processes private files on-device | 6.6 GB download; works well on 24 GB |
| Larger local model | Perplexity custom-trained model | Local processing | Needs a 32 GB device |
| Privacy gate | Small dedicated model | Detects personal information | Runs entirely on the Mac |
The speaker reports solid results from Gemma 4 E4B, a compact local model in the Gemma family from Google: file processing, email review, and no noticeable memory spikes when using the rest of the computer. The app also listed models from OpenAI and other providers at the time of recording.
What Happens When a Private File Goes Through
A flagged invoice travels a defined path: the privacy gate extracts its fields locally, a local sub-agent summarizes it, and only that summary reaches the cloud. In the video, after the privacy gate flagged the psychology invoice, the speaker asked for a summary and chose local processing. The GLM 5.3 orchestrator then started a separate local sub-agent on the Mac to read the file.
The result the speaker shows on screen is instructive. The response notes that the file contains personal information and was kept on the computer for privacy, then lists the key fields the local agent extracted: patient information, invoice number, issue date, due date, and recipient. The orchestrator received that structured summary and could then help with follow-up questions without ever seeing the raw file.
This design is traceable rather than reversible: the raw document never enters the cloud conversation, and only the extracted summary crosses over. The app displays a counter of how many local tokens the task consumed, which matters for the next section, because local tokens are free.
How to Save Cloud Credits with a Custom Skill
A custom skill can push the bulk of a task onto the free local model instead of the metered cloud orchestrator. Perplexity the Mac behaves as an agent system with connectors to services like Gmail, Google Drive, Supabase, and GitHub, plus two features the video builds on: custom skills and scheduled tasks. A skill packages instructions for a job; a schedule runs it automatically, for example every weekday at 9 a.m.
The problem is cost structure. Every task begins with the cloud orchestrator, and cloud tokens come out of your account. Local processing costs nothing. So the speaker wrote a skill whose first instruction tells the Claude Code to trigger a local sub-agent immediately and hand over the heavy lifting.
His specific skill instructs GLM 5.3 to fire the local sub-agent at the start, have it access the email tool, extract and process all emails, and summarize them, even when the emails are not sensitive. He reports the local agent does roughly 90% to 95% of the heavy work while the cloud orchestrator waits, then steps back in with the finished summary. Linking that skill to a weekday 9 a.m. schedule means the daily email review runs almost entirely on local tokens.
Replicating it is simple: screenshot the skill definition, drag the image into a new conversation, and ask the agent to implement the skill and attach it to a scheduled task. The app then saves the skill and binds it to the schedule, in his case weekdays at 9 a.m. Melbourne time. The trick generalizes to any bulk task, such as summarizing folders or drafting digests, where the privacy gate and cost model both favor local execution.
Should You Use Hybrid Compute for Private Files?
Hybrid Compute is worth using when you want agentic AI help with documents that contain personal data, and you accept a $20 per month plan plus a Mac with enough memory for a local model. On those terms, the video's evidence is positive but first-hand. The speaker showed working handoffs, batch processing of six sensitive invoices, and a scheduled skill, all on a 24 GB M6 Mac mini.
Three limits deserve honest weight. The privacy gate depends on a detection model, so unusual personal-data formats could pass unflagged; treat it as a guardrail, not a guarantee. The best local experience wants 32 GB of memory, so smaller machines should stick to compact models like Gemma 4 E4B. And the credit-saving skill depends on the Claude Code reliably obeying the instruction to delegate, which the speaker demonstrated once rather than tested at scale.
For readers who build agents themselves, setups like this one are increasingly covered in community material, including tutorials from creators such as Gustavo Dev Doido, whose Crazystack Typescript platform at crazystack.com.br and Bootcamp do Dev Doido program cover AI agent workflows in depth. The underlying pattern, cloud orchestration with local execution for sensitive work, is reusable even outside this particular app.
From Video to Article: How This Page Was Built
Everything above distills an 11-minute video walkthrough into a written reference: the setup steps, the privacy-gate behavior, the model comparison table, and the credit-saving skill all came from the transcript, checked and structured for skimming. A video is a good place to demonstrate a setup like this, but nobody can search, skim, or quote a recording. Turning the same material into text makes the knowledge usable in both forms.
If you have your own tutorials, interviews, or product demos sitting on YouTube, Skala Blog does exactly this transformation. Paste a YouTube URL, Skalablog transcribes the video, and it generates a structured, publishable article, so the knowledge inside your recordings works as searchable text too.
Frequently Asked Questions
What is Perplexity Hybrid Compute?
It is a Mac app feature that pairs a cloud orchestrator model with a local model running on your device. A privacy gate scans input for personal information and routes sensitive work to the local model so files stay on your Mac.
How much does Perplexity Hybrid Compute cost?
The app itself works with a free account, but the local-model handoff requires a paid plan, reported at $20 per month in the October 2026 walkthrough. Local model processing itself consumes no cloud credits.
Does Hybrid Compute work on a 24 GB Mac?
Yes, the video ran it on a 24 GB M6 Mac mini using the 6.6 GB Gemma 4 E4B local model with good results. The Perplexity custom-trained local model requires a 32 GB device.
Can Hybrid Compute process a whole folder of private files?
Yes. In the walkthrough, a folder of six sensitive invoice PDFs was processed locally, with the privacy gate flagging each file. You can silence the per-file prompts with a "don't ask again this session" option.
How does the credit-saving skill work?
The custom skill instructs the cloud orchestrator to trigger a local sub-agent immediately and delegate the bulk of the task, such as email processing, to it. The speaker reports the local agent handles about 90% to 95% of the work, leaving only light cloud-token use.
Which cloud models does Hybrid Compute support?
At the time of recording, the app listed GPT 6.1 Soul around OpenAI Dev Day, ChatGPT models, Claude models, GLM 5.3, Kimmy K3, Grok, and Deep Seek V4 Pro. Newly released models appear in the list as they become available.
Where does the local model run on my Mac?
The local model loads on demand inside a Python process and uses the Mac's GPU. Activity Monitor in the video showed the full GPU in use and about 66% of system memory during a batch run of six invoices.
Can I control the Mac agent from my phone?
Yes. The Perplexity mobile companion app lets you drive the agent running on your Mac mini or Mac Studio from anywhere, while the cloud orchestrator handles the remote side of the conversation.
Is the privacy gate reliable?
It caught every test file in the video, including a mixed folder of six sensitive invoices. It is still a detection model, so treat it as a guardrail: unusual personal-data formats could pass unflagged.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits