Can a 32GB Mac Run Perplexity Computer Local? 19.4 GB at 4-bit

A 32GB Mac mini M6 cannot run Perplexity Computer in fully local mode. Perplexity recommends 36 GB of unified memory or more for that mode, and Apple caps the standard M6 Mac mini at 32 GB. The same machine can still run Hybrid Compute, which Perplexity's help center says needs 24 GB or more and gives its best results at 32 GB. Both thresholds come from Perplexity's own pages, and both decide whether the Mac you own, or the one in your cart, can run Computer at all.

TL;DR: Perplexity shipped Computer fully local on Mac on October 20, 2026, and fully local asks for 36 GB of unified memory or more. Hybrid Compute runs from 24 GB, with 32 GB for best results. The M6 Mac mini tops out at 32 GB, so the cheapest machines that clear the fully local number are the Mac mini M5 Pro at 48 GB or 64 GB and the Mac Studio M5 Max at its 36 GB base. The 19.4 GB is Perplexity's own size for the 4-bit Qwen3.6-35B-A3B checkpoint, not the size of the app's download.

The short answer: 36 GB for fully local, 24 GB for Hybrid

Perplexity states the fully local threshold in two places, in the same words. The launch post says that "for running Computer locally on Mac, 36 GB or more is recommended". The product page puts it in a requirements block: fully local, 36 GB unified memory recommended.

Hybrid Compute is the lower bar, and it is documented separately:

• The Perplexity help center lists an Apple silicon Mac with 24 GB or more of unified memory on macOS 15 Sequoia or later, in the local mode column labelled Hybrid compute. The same page adds that on a Mac, 32 GB of unified memory gives the best results.
• The product page's requirements block lists 32 GB unified memory recommended for Hybrid.

So the honest reading of the two official pages is: Hybrid starts at 24 GB and is recommended at 32 GB. Fully local starts at 36 GB. Those are not the same number, and merging them would be wrong in both directions.

Hybrid Compute is Perplexity's split mode: cloud models handle research and reasoning while a model on your Mac reads private files and flags personal information before anything leaves the device. Fully local puts the model, the agent that directs the task, and the task execution all on the Mac, with no cloud models in the loop.

Both modes need a paid plan. Local Computer on Mac is available to Perplexity Pro, Max and Enterprise subscribers, and the local models list carries the same restriction. There is no free tier for either mode.

Which Macs qualify: the memory table

Apple's own configuration options decide this, not Perplexity's supported-chip list. Perplexity only says "supported Apple silicon Macs, like the new Mac mini with M6 and M5 Pro", which is why the memory number does the work.

The memory options in the table come from Apple's Mac mini technical specifications and Mac Studio technical specifications, read on October 21, 2026. Apple's September 22 newsroom post introduces the same two desktops and their M6, M5 Pro, M5 Max and M5 Ultra tiers.

Three consequences follow.

The standard Mac mini M6 is out of reach for fully local, at every configuration. Its ceiling is 32 GB, four gigabytes short, and that is the whole verdict for the most popular cheap Mac desktop of this cycle. It is a fine Hybrid machine.

The mini line clears 36 GB only at an upgrade. A Mac mini M6 never clears it, and a Mac mini M5 Pro does so at 48 GB or 64 GB, not at its 24 GB base. The Mac Studio M5 Max is the one Apple desktop that clears 36 GB at its base configuration, with the 36 GB tier documented in our Mac Studio M5 Max 36GB guide.

A laptop qualifies too. The machine Perplexity benchmarked its engine on is a MacBook Pro with an M5 Max and 128 GB of unified memory, and any Apple silicon laptop configured at 36 GB or more meets the recommendation. AppleInsider notes that older Apple silicon Macs with enough memory also qualify, since Perplexity frames 36 GB as a recommendation rather than a hard minimum.

If your Mac tops out at 32 GB and you want a local model to run today, that machine is still one of the better local LLM hosts in its class. Our Mac mini M6 32GB picks cover what actually fits it, and the M5 Pro 64GB guide covers the tier that clears Computer's fully local bar.

Where the 19.4 GB comes from

Perplexity's Apple Silicon engineering post describes the checkpoint its local engine runs. Groupwise affine 4-bit quantization is the scheme: every weight is a 4-bit integer code, and every group of 64 weights shares one bfloat16 scale and bias. In Perplexity's words, that "reduces the 35-billion-parameter model from roughly 70 GB of bfloat16 weights to a 19.4 GB checkpoint".

That number describes one artifact, and it is worth being precise about which one. Perplexity does not say which file the Mac app downloads, and the 19.4 GB belongs to the Qwen3.6-35B-A3B checkpoint used by Lily, the local inference engine in that post. It is not a published download size for Perplexity Computer, and reading it as "the app is 19.4 GB" is a mistake.

The model Perplexity calls PPLX Pro is also not published under that name. There is no PPLX Pro repository on Hugging Face, so there is no downloadable size for it and no way to check its weights. What the perplexity-ai org does publish is the surrounding family: decider models, embedding models, PII Tracer, and a set of public pplx-computer-qwen builds, including an MLX 4-bit build of Qwen3.6-35B-A3B in four safetensors shards, an MLX 4-bit build of a Qwen3.8-27B checkpoint, and a llama.cpp GGUF of the same 27B that ships with llama-server commands in its model card.

The app's local model menu is the only place the Mac models are named together. The launch post says Computer "runs Qwen 3.6 or PPLX Pro, a post-trained version of Qwen". The product page's setup step says to download either PPLX Pro or Qwen 3.6. The help center lists three local models for an Apple silicon Mac: Gemma 4 E4B, Qwen3.6 35B-A3B, and the Perplexity model. For the model itself, see our Qwen 3.6 on Mac guide.

What 36 GB buys beyond the checkpoint

Perplexity's own measurements of the 27B build show where the memory goes. On a MacBook Pro with 36 GB, generating a 118-token code answer greedily, the MLX 27B model card reports a peak of 16.4 GB with plain generation and 17.1 GB with the MTP speculative drafter attached. The quantized weights alone are described as about 15 GB at 4.695 bits per weight.

That is the shape of the budget. A 27B-class local model at 4-bit lands around 16 GB of resident memory, and the 19.4 GB checkpoint for the larger mixture-of-experts model is bigger still. On the 19.4 GB checkpoint, a 36 GB Mac is left with a little over 16 GB for macOS, the Perplexity app, the sandbox that runs code and tools, the KV cache that grows with context, and whatever you have open while a task runs. Our arithmetic, not Perplexity's, and it explains a recommendation that looks high next to the checkpoint alone.

A mixture of experts (MoE) model is a design that holds all of its parameters in memory but activates only a fraction per token: Qwen3.6-35B-A3B has 35 billion parameters and activates about 3 billion of them. Memory is sized by the total, speed by the active part.

The same 36 GB machine is also fast enough to be useful. Perplexity's model card reports 22.4 tokens per second for plain generation on that MacBook Pro, rising to 48.0 tokens per second with the MTP drafter attached, at 87% draft acceptance. Those are single-model generation numbers from the vendor, not Computer task times.

The speed numbers are Perplexity's, measured on one Mac

The engineering post reports engine throughput for Qwen3.6-35B-A3B in 4-bit, on a single MacBook Pro with an M5 Max, a 40-core GPU and 128 GB of unified memory. At a 4K-token prompt and a 4K-token decode context, the engine reaches 5,749.9 prefill tokens per second and 186.6 decode tokens per second, against 4,737.5 and 140.9 for MLX-LM at the same settings. Across ten prompt lengths for prefill and ten context lengths for decode, from 256 to 128K tokens, Perplexity reports averages of 1.22 times MLX-LM's prefill throughput and 1.32 times its decode throughput.

Those are engine numbers from one vendor, on one machine, and they are not Computer task times. Nothing published by Perplexity describes how a full agent workflow behaves on a 48 GB mini, a 36 GB Mac Studio, or a MacBook Pro with less memory, and there is no independent Mac reproduction of these figures at all. The engine itself is open source: Lily is a Rust and Metal server for one checkpoint, exposed as an OpenAI-compatible chat-completions endpoint, and its README requires an Apple GPU family 10 or later, which means M5 and newer, for the standalone demo.

Fully local is not the same as offline-only

The distinction matters for what you can actually ask Computer to do on a plane.

Perplexity's product page states it plainly: "Fully local describes where the AI runs. Sending a message or meeting invitation still uses that app's service." A local task that reads your Notes, updates a workbook and drafts a reply keeps the model and the file work on the Mac. The moment a step has to send that email, post to a connected app, or look something up on the web, the network is back in the path, and the app asks for permission before escalating. The launch post says the local orchestrator can call on one of 15+ frontier models for advanced reasoning when you authorise it.

The privacy claim is therefore about processing, not connectivity: your files and prompts are not sent to cloud models in fully local mode, while outbound actions still leave through the app they belong to.

What is not verified

• Which model file the Mac app downloads, and how large it is. Perplexity publishes no download size for PPLX Pro or for the Qwen 3.6 option in the app.
• Any independent Mac measurement of Perplexity Computer or of Lily. The throughput figures above are vendor numbers on one M5 Max with 128 GB, and no lab, reviewer or benchmark suite has reproduced them.
• Whether 36 GB is enforced in code. Perplexity calls it a recommendation, so a machine below it may still start and simply swap or refuse long contexts. No published error message or memory check exists.
• Whether Hybrid starts below 24 GB. The help center says 24 GB or more, and MacRumors reported Hybrid "works on Macs with as little as 24GB", so the two agree. Nothing published says what happens at 16 GB.
• What PPLX Pro actually is. The name appears on Perplexity's pages, but no repository, card or size for it exists anywhere public.

FAQ

Can a 32GB Mac mini M6 run Perplexity Computer fully local?

No. Perplexity recommends 36 GB or more of unified memory for fully local operation, and Apple's specification page caps the M6 Mac mini at 32 GB. That machine can run Hybrid Compute instead, where the cloud does the research and reasoning and your Mac handles private files locally from a 24 GB floor, with 32 GB giving the best results.

Is 19.4 GB the size of the Perplexity Computer download?

No. The 19.4 GB is Perplexity's own figure for the 4-bit Qwen3.6-35B-A3B checkpoint described in its Apple Silicon engineering post, down from roughly 70 GB in bfloat16. Perplexity does not state which file the app fetches or how large the download is, so treat any number presented as the app size as unverified.

Does fully local mean Computer works with no internet connection?

Only partly. The model, the agent and the file processing run on the Mac, so private work does not depend on a network. Steps that send a message, post to a connected app or fetch current information still use those services, which is why Perplexity says fully local describes where the AI runs and not whether the task stays offline.

Can I run PPLX Pro outside the Perplexity app?

Not from a public download. There is no PPLX Pro repository on Hugging Face, so its weights and size are unknown. Perplexity does publish Qwen-based Computer builds you can load yourself with MLX or llama.cpp, including a 4-bit Qwen3.6-35B-A3B MLX build and a Qwen3.8-27B GGUF with speculative drafters, but those are model files rather than the Computer agent.

Do I need a paid plan for either mode?

Yes. Local Computer on Mac is available to Perplexity Pro, Max and Enterprise subscribers, and local models carry the same restriction. Local work does not consume Computer credits, while any step that runs in the cloud does, which is the trade the fully local option is designed to make explicit.

Sources

• Perplexity, Computer now runs fully local on Mac, October 20, 2026 (36 GB or more recommended, model options, subscriber tiers).
• Perplexity, Run Computer tasks locally on your Mac, product page (36 GB fully local, 32 GB Hybrid, offline caveat, setup steps).
• Perplexity Help Center, What is Personal Computer? (24 GB or more for Hybrid compute on Apple silicon, 32 GB best results, local model list).
• Perplexity Engineering, Optimizing on-device inference for Apple Silicon, September 1, 2026. The live page blocks automated fetches, so the figures quoted above were read from the archived snapshot of October 16, 2026.
• Hugging Face, perplexity-ai/pplx-computer-qwen-3-8-27b-mlx-20260824 (peak memory and generation figures on a 36 GB MacBook Pro).
• Hugging Face, perplexity-ai/pplx-computer-qwen-3-6-35b-a3b-mlx-20260709 and perplexity-ai/pplx-computer-qwen-3-8-27b-dflash2-gguf-20260826.
• GitHub, Perplexity Lily inference engine (supported hardware, one checkpoint, OpenAI-compatible API).
• Apple, Mac mini technical specifications and Mac Studio technical specifications, read October 21, 2026, plus the September 22, 2026 newsroom post.
• AppleInsider, Perplexity Computer can run completely local on Mac, with a big caveat, and MacRumors, Perplexity Computer Can Now Run Fully Locally on Mac, for press coverage of the 24 GB Hybrid figure.

Every number here is a vendor figure opened at the source, an Apple specification, or our own arithmetic labelled as such. For the same 36 GB question on a different engine, our Splash Engine requirements breakdown covers Inco's floor. How we work: about ModelFit. Corrections welcome: contact.