Perplexity Hybrid Compute, launched September 1, 2026 in the Perplexity Mac app for Pro ($20 a month) users, is the better default for founders who want research plus privacy without setup: an on-device classifier swaps names, addresses and account numbers for stand-ins before anything reaches the cloud. LM Studio Bionic 1.1, free with local models, wins when the whole task must stay on the machine. Radically Human's IDEAS framework decides.

The question I get from founders about private AI has changed shape this year. In 2025 it was "can a local model do anything useful." In September 2026 it is "which of these two very different privacy designs should I trust with the cap table," because the two products that matter make opposite bets.

Perplexity's bet is that you should not have to choose. Hybrid Compute, which launched September 1, 2026 as part of the Perplexity Mac app, lets its Computer agent start a task in the cloud with frontier models, then hand the steps that touch private files or sensitive data to a local model on your Mac. An on-device PII classifier, which Perplexity has open-sourced, reads each task before it leaves the machine and swaps names, addresses and account numbers for stand-ins, restoring them when the answer returns. The local model is PPLX Qwen 3.8 27B, downloaded in one click with no Ollama or runtime setup, and work it handles uses no cloud credits. The requirements are the catch: Apple silicon, macOS 15 or later, 24 GB of unified memory minimum with 32 GB recommended, so 8 GB and 16 GB Macs are out. It sits inside paid plans, Pro at $20 a month or Max at $200. LM Studio's bet is the opposite: keep everything on the machine unless you deliberately reach out. Bionic, the agent app LM Studio introduced in July 2026 as a separate product from the classic model runner, runs projects and parallel sessions that edit documents, spreadsheets and PDFs, write and run code, and transcribe voice offline, each session against a local model, a model on another device through LM Link, or a frontier open model such as GLM 5.2, Kimi K3 or DeepSeek V4 Pro in LM Studio's Secure Cloud under a Zero Data Retention policy. The 1.1.0 release in September 2026 merged the Code and Work project types into one. The app, local models, offline transcription and LM Link for up to five devices are listed at $0; cloud models bill per million tokens from purchased credits, and an announced Bionic Pass is not yet purchasable. It downloads for Apple silicon from lmstudio.ai and supports reusable Skills in the open SKILL.md format.

To judge them I used Paul Daugherty and H. James Wilson's Radically Human, because their IDEAS framework was written for exactly this trade-off: Intelligence (how much reasoning the job needs), Data (how little of it should move), Expertise (whether humans can teach the system their way of working), Architecture (whether the system is a living one that can shift between modes) and Strategy, with their separate argument that trust is now a competitive asset rather than a compliance box. Turned into a working procedure for one founder and one confidential task, it becomes what I call the Sensitivity Split:

  1. Classify by what leaves the machine, not by how secret the topic feels. List the actual strings in the task: names, figures, account numbers, unreleased numbers. Done when you can say which of them may transit a cloud API and which may not.
  2. Match the intelligence to the step, not the task. Reading a 40-page contract and pulling clauses is a 27B-local job; reasoning about whether to sign it is a frontier job. Done when each step has a model tier next to it.
  3. Decide who does the classifying. Perplexity's classifier catches PII patterns; it does not know that "Project Heron" is your acquisition target. If the sensitivity is contextual rather than pattern-shaped, you must do the split by hand, which favors Bionic's all-local session. Done when you have named the classifier, human or software.
  4. Teach the workflow once. Daugherty and Wilson's machine-teaching point: write the recurring task down. In Bionic that is a SKILL.md; in Perplexity it is a saved Computer task. Done when a second run needs no re-explaining.
  5. Check the blast radius before the first run. Bionic edits files dragged in from outside a project in their original locations, so work on copies; Perplexity's local steps use no credits but its cloud steps do. Done when a wrong answer costs you a file copy, not the file.

Run through that, my verdict for the typical founder, who already pays for one research tool and wants a board deck drafted from confidential numbers, is Hybrid Compute. The Architecture row is decisive: it is the first mainstream product where the split between local and cloud happens inside one task rather than in the user's head, and the Data row follows, because the classifier moves less raw information than most people manage manually. Bionic wins the two cases where that design is the wrong shape: when the sensitivity is contextual and no PII pattern would catch it, and when the budget is zero, since a Bionic session against a local model costs nothing and never opens a socket.

The failure modes are what I would want to know before committing either to real documents. Hybrid Compute's classifier is a pattern detector; a strategy memo with no names in it can be entirely sensitive and entirely cloud-eligible. A 27B local model is competent at extraction and weak at multi-step judgment, and both products lean on it for the private steps. Perplexity's design sends part of every task to the cloud by construction, and the credit meter on Computer tasks is real money on Pro. Bionic's Secure Cloud is zero-retention but still a cloud, its in-place file editing is a genuine trap, and its platform support has shifted month to month, so verify the download page rather than any review, including this one. Both products are weeks or a few months old. Daugherty and Wilson's phrase for the right posture is Forever Beta; treat your setup that way, and re-run the Sensitivity Split whenever either vendor changes what leaves the machine.

Sources: 9to5Mac, "Perplexity launches privacy-minded 'hybrid compute' AI feature for Mac," September 1, 2026 (9to5mac.com); LM Studio, "Welcome to LM Studio Bionic" documentation and the Bionic product page (lmstudio.ai/docs/bionic, lmstudio.ai); The Rundown AI, "LM Studio Bionic Review: Pricing, Local AI & Safety," pricing checked August 31, 2026 (therundown.ai/tools/lm-bionic); GyanAangan, "LM Studio Bionic vs Classic LM Studio," updated September 14, 2026 with the 1.1.0 note (gyanaangan.in); GEO Toolbox, "Perplexity Pricing 2026," updated September 4, 2026 (geotoolbox.ai); Paul Daugherty and H. James Wilson, Radically Human (Harvard Business Review Press, 2022).

Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · What University Will Not Teach You