By Billy G.R. · 2 October 2026
Checked 2026-10-02. Author: Billy G.R. Retail prices move; the hardware catalog stores the Amazon snapshot, not a promise of stock.
Claim. Perplexity Comet is Perplexity’s AI browser. This guide does not restate a feature matrix, because a fetch of perplexity.ai/comet timed out on 2 October 2026.
Method. Primary URL recorded and attempted. Local recommendations below do not depend on unpublished Comet features.
As of.
Claim. A text-only local browser agent can sit on a 7–14B Q4 model (6.2–10.4 GB). A screenshot agent wants a 27–30B multimodal model, about 18–20 GB.
Method. Site formula at Q4_K_M. browser-use repository confirmed at github.com/browser-use/browser-use.
As of.
Claim. The local stack does not include Perplexity’s web index. It automates a browser you run.
Method. Product distinction, not a benchmark. Privacy and offline use are the reason to switch, not search quality.
As of.
Methodology — parameter math, quantization bytes, and the source list.
Perplexity Comet is Perplexity’s browser. Enterprise notes live at the help center and the enterprise post. I tried to re-read the product page on 2 October 2026 and the fetch timed out, so I am not going to invent a sidebar checklist from memory. Treat capabilities, permissions, and plan gates as whatever that page says when you open it. What this guide can say without that page: Comet is their product, it is cloud-assisted, and a local browser agent will not match Perplexity’s search index. If you need their index, use Comet. If you need the click-path to stay on a machine you control, build the stack below.
browser-use is an open-source library for driving a browser from a language model. The GitHub description, confirmed the same day, is “Agents that use the browser.” You pair it with Playwright and Chromium, and with a local model served by Ollama. Open WebUI stays the chat door for everything that is not a browser task. OpenHands can wrap a browser tool; I am not linking their old SDK guide path, because the docs index has since split into Agent Canvas and the Software Agent SDK and I did not re-confirm that deep link.
The model sees either the page text you extract or a screenshot. Those are different memory problems. Text-only browsing can live on a small model. Vision browsing wants a model that accepts images, and those weights are larger. The browser-use hardware page goes further on reliability. This page is the product comparison.
Site formula, Q4_K_M: params × 0.6 + 2. Chromium itself wants system RAM. On a unified-memory Mac that RAM is the same pool as the model. On a discrete GPU the model uses VRAM and Chromium uses the host. The calculator does not know about Chromium. Leave headroom.
| Agent style | Model | Memory |
|---|---|---|
| Text extracted from the page | K2-Horizon-7B or a 14B Q4 | 6.2 GB or 10.4 GB. A 12 GB card can try it. Do not add a second model. |
| Screenshots, charts, visual layout | Qwen3.8-27B if the build is multimodal, or Muse Glimmer 30B | 18.2 GB or ~20 GB formula. Plan 24 GB, not 16 GB. |
| Browser plus a separate chat model | Do not load both at full size | Share one server. Two resident 27B models are ~36 GB of weights before KV cache. |
I am not quoting a success rate for browser-use. Page changes break selectors whether the model is local or not. The hardware question is only whether the model and the browser fit at the same time.
If the machine’s job is “sit there and drive a browser overnight,” a Mac mini M5 Pro 24 GB is the small always-on option in the verified set: ASIN B0HGGHNQY6, $1,669.99 on 1 October 2026. It is enough for text-mode 7B and tight for a 27B vision loop. I would not step up to the EVO-X2 just to run Comet’s local cousin unless you also want a large MoE on the same box. Comet itself, if you keep it, needs no VRAM at all.
Mac mini M5 Pro 24 GB on Amazon
ASIN B0HGGHNQY6. Amazon Associates link. The live price is on that page, not locked in here.
A correction or a hardware question goes to the contact page.