By LLM Hardware team · Methodology ·
Checked 2026-10-13. Author: LLM Hardware team. This site does not store or show Amazon prices. Check current price on Amazon.
Key fact: Meta Muse, ChatGPT dots, Grok Bot, Manus, Claude Cowork, Perplexity Comet and Gemini skills run on vendor cloud computers. None of them needs local VRAM.
How we got it: Each vendor page linked in the sections below, read 2 October 2026 and rechecked 14 October 2026.
Checked:
Key fact: A local replacement brain needs about 6.2 GB at 7B, 10.4 GB at 14B, 18.2 GB at 27B and 20 GB at 30B, all at Q4.
How we got it: VRAM formula: params x 0.5 x 1.2 + 2. Tool, browser and container memory are extra.
Checked:
Key fact: Speed figures are not published on this page.
How we got it: Fit comes first; tokens per second depend on the runtime and context length.
Checked:
Methodology — parameter math, quantization bytes, and the source list.
Cloud agents from Meta, OpenAI, xAI, Manus, Anthropic, Perplexity and Google all run on the vendor's computers. Replacing one with a local agent moves the model onto your hardware, so the question becomes how much VRAM or unified memory the replacement model needs. This page covers all seven products in one place, with the local stack and the memory tier for each.
| Memory | Largest agent model at Q4 | VRAM formula |
|---|---|---|
| 8 GB | 7-8B | 6.2-6.8 GB |
| 12-16 GB | 14B | 10.4 GB |
| 24 GB | 27-32B | 18.2-21.2 GB |
| 48 GB+ or 64 GB unified | 70B | 44 GB |
Check your exact context length in the VRAM calculator and read the agent-workload VRAM guide before adding a browser or container on top.
What it is (Meta, Introducing Muse (8 Sep 2026)): a consumer personal agent that runs on a dedicated cloud Muse Secure VM, powered by Muse Spark, messaged from the Muse app or WhatsApp. It is not a model you load on a GPU. Muse Glimmer, the 30B open-weight model, is a separate product.
Local replacement: Ollama or LM Studio for the weights, Open WebUI as the always-on front door, and OpenHands or browser-use for actions. Muse Glimmer 30B is the open model if you want Meta weights.
Memory tier: 7B on 8-12 GB (6.2 GB). Muse Glimmer 30B at Q4 is 20 GB, so 24 GB.
Source: Meta, Introducing Muse (8 Sep 2026). Older detail page: Meta Muse vs local.
What it is (ChatGPT features, dots (read 2 Oct 2026)): always-on ChatGPT agents, each with a cloud computer. The feature page lists Pro, Business Premium and Enterprise access; messaging is through ChatGPT, Slack and Teams.
Local replacement: a 27-30B model behind Open WebUI with scheduled tasks and tools you enable.
Memory tier: 27B at Q4 is 18.2 GB and 30B is 20 GB before tool overhead, so a 24 GB GPU or a 36 GB Mac.
Source: ChatGPT features, dots (read 2 Oct 2026). Older detail page: ChatGPT dots vs local.
What it is (xAI, Introducing Grok Bot (read 2 Oct 2026)): xAI's agent product. The work runs on a cloud computer the bots share; the desktop and iOS apps are control surfaces. Access is through SuperGrok and Cursor plans per the post.
Local replacement: a local teammate built from Ollama plus CrewAI or LangGraph for multi-step work.
Memory tier: 7B helper at 6.2 GB; 27B teammate at 18.2 GB (24 GB GPU).
Source: xAI, Introducing Grok Bot (read 2 Oct 2026). Older detail page: Grok Bot vs local.
What it is (manus.im (fetched 2 Oct 2026)): a hosted cloud-sandbox agent. The homepage did not publish a feature matrix that could be quoted, so no feature claims are made here.
Local replacement: OpenHands (Agent Canvas, CLI or SDK) in a local container with a coding model.
Memory tier: 27B coding model at Q4 is 18.2 GB; 24 GB is the minimum comfortable tier, 36 GB leaves KV-cache headroom. Container RAM is extra.
Source: manus.im (fetched 2 Oct 2026). Older detail page: Manus vs local.
What it is (Anthropic, Claude Cowork product page (read 2 Oct 2026)): Anthropic's task mode beside Chat. It works in folders and tools you choose, can run scheduled tasks in the cloud, and its computer use is described as a research preview.
Local replacement: LM Studio desktop agents or Open WebUI with a mounted folder and approval prompts.
Memory tier: 27B document agent at 18.2 GB; a 30B multimodal model for screenshots is 20 GB.
Source: Anthropic, Claude Cowork product page (read 2 Oct 2026). Older detail page: Claude Cowork vs local.
What it is (Perplexity Comet): Perplexity's AI browser. Its page blocked automated fetches on 2 Oct and 14 Oct 2026, so no feature claims are restated. The local stack does not include Perplexity's web index.
Local replacement: browser-use driving a browser you run.
Memory tier: Text-only browser agent on 7-14B (6.2-10.4 GB); screenshot agent on 27-30B multimodal (18-20 GB).
Source: Perplexity Comet. Older detail page: Perplexity Comet vs local.
What it is (Google, skills in Gemini (30 Sep 2026)): saved, reusable task instructions in Gemini. Google's post says skills replace Gems, with Gems support removed starting in November for personal accounts and in 2027 for Workspace.
Local replacement: Ollama Modelfiles or Open WebUI presets: a saved system prompt plus tools.
Memory tier: 7-14B (6.2-10.4 GB) for a saved prompt; 27B (18.2 GB) for more reliable tool use.
Source: Google, skills in Gemini (30 Sep 2026). Older detail page: Gemini skills (formerly Gems) vs local.
Local wins when files must stay on hardware you own, when you want to work offline, or when per-seat subscriptions add up across a team.
A 24 GB Mac mini holds a 7-14B agent model comfortably. A 128 GB Strix Halo mini PC holds 70B dense or large MoE models. Check current price on Amazon.
Mac mini M5 Pro 24 GB on Amazon
ASIN B0HGGHNQY6. Amazon Associates link. The live price is on that page, not locked in here.
GMKtec EVO-X2 128 GB on Amazon
ASIN B0F53MLYQ6. Amazon Associates link. The live price is on that page, not locked in here.
No. Meta Muse, ChatGPT dots, Grok Bot, Manus, Claude Cowork, Perplexity Comet and Gemini skills run on the vendor cloud. You only need local VRAM if you replace them with a local model.
A 7B agent model at Q4 needs about 6.2 GB, so an 8 GB GPU. A 27B model needs about 18.2 GB, so a 24 GB GPU or a 36 GB Mac. Add headroom for the browser or container.
A correction or a hardware question goes to the contact page.