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. Meta Muse, ChatGPT dots, and Grok Bot are three cloud always-on agents. None of them is a weight file you load locally.
Method. Meta announcement 8 September 2026 (Muse Secure VM, Muse Spark). chatgpt.com/features/dots/ (GPT-6 Astra, cloud computer, Pro / Business Premium / Enterprise). x.ai/news/introducing-grok-bot (shared cloud computer, desktop and iOS as control surfaces). All read 2 October 2026.
As of.
Claim. The local alternative is one model server plus a UI. A 27B Q4 needs about 18.2 GB. A 7B Q4 needs about 6.2 GB. A 70B Q4 needs about 44 GB.
Method. Site formula params × 0.5 × 1.2 + 2. Qwen3.8-27B, K2-Horizon-7B, and a generic 70B dense count.
As of.
Claim. Calling the local stack “compatible with Muse” would be false. It is an alternative you run yourself.
Method. Different vendors, different computers, no shared protocol claimed on the primary pages.
As of.
Methodology — parameter math, quantization bytes, and the source list.
Three products are soaking up the “always-on personal agent” searches, and they are easy to smear together because each one messages you and each one has a computer that is not yours. Meta Muse runs on a Muse Secure VM and is powered by Muse Spark. ChatGPT dots run on GPT-6 Astra, each with a cloud computer. Grok Bot gives named teammates a shared cloud computer; the writeup is at x.ai/news/introducing-grok-bot. This page is the comparison. The local stack at the bottom is an alternative. It is not “Muse compatible,” “a dot,” or “Grok, self-hosted.”
| Meta Muse | dots | Grok Bot | |
|---|---|---|---|
| Where it runs | Cloud VM. Sentinel gates network access. | Its own cloud computer. You can also connect yours. | Shared cloud computer. Desktop and iOS are the remote controls. |
| Model named | Muse Spark. Not Glimmer. | GPT-6 Astra. | Not published as a downloadable weight. |
| How you talk | Muse app or WhatsApp. iOS, Android, muse.ai in the US rollout described. | ChatGPT, Slack, Teams. Texting listed as coming next. | Desktop and iOS apps. Also via paid Cursor plans. |
| Who can start | Free for most uses, subscriptions for more, per Meta’s post. | Pro, Business Premium, Enterprise (admin). Not described as free. | SuperGrok tiers and listed Cursor paid plans. Enterprise waitlist on the news post. |
| Local VRAM | None. | None. | None. |
Longer writeups: Muse, dots, Grok Bot. Operator is not a fourth column. It is not a live product on the dots page. Muse Glimmer is not a fourth cloud agent. It is the 30B open model, covered on its own page.
Pick a single brain and put a door in front of it.
That stack will not book travel through Meta’s checkout, will not see your ChatGPT memory, and will not sign into Grok’s cloud computer. It will run when those services are down, on files you mounted. Whether that trade is worth it depends on the chore, not on a benchmark I am not going to invent.
Formula: VRAM ≈ params × bytes × 1.2 + 2. Q4 bytes are 0.5, so Q4 is about params × 0.6 + 2. Q8 is params × 1.2 + 2. MoE memory follows total parameters, not active ones. Confirm a context length on the VRAM calculator. The scaffold pad of 1–4 GB is explained in the agent formula guide and is not part of the formula.
| Tier | Iron we can name | Comfortable Q4 dense |
|---|---|---|
| 6–12 GB | Laptop iGPU, RTX 3060 12 GB, RTX 5070 12 GB (B0DS6S98ZF) | 7B = 6.2 GB. 14B = 10.4 GB, tight. |
| 16–24 GB | RTX 4090 24 GB (B0BG94PS2F). Mac mini M5 Pro 24 GB (B0HGGHNQY6). | 27B = 18.2 GB. 32B = 21.2 GB, tight. Glimmer Q4 formula 20 GB. Meta K-quant ~17 GB. |
| 36–64 GB | Mac Studio M5 Max 36 GB (B0HGKSQMX6). 64 GB Studio configures at Apple. | 32B comfortable. 70B Q4 = 44 GB needs 48 GB or more. |
| 96–128 GB | GMKtec EVO-X2 128 GB (B0F53MLYQ6). ~256 GB/s class. No CUDA. | Large MoE aggressive quants. DeepSeek-V4-Flash is 284B total. |
| 192 GB+ | Framework Desktop 192 GB, Mac Studio Ultra. Apple, not an invented ASIN. | GLM-5.3-Flash and MiniMax-M3 class MoEs. Not a single gaming GPU. |
On 1 October 2026 Amazon had the mini at $1,669.99, the Studio at $2,449, and the EVO-X2 at $3,649.99. Those are snapshots. The mini is the 7B always-on box and a tight 27B. The Studio is the daily 27–32B agent. The EVO-X2 is the MoE box. An RTX 5090 at 32 GB (B0DS2X13PH) showed $7,398 on 2 October 2026. It still does not hold a 70B Q4 at 44 GB. I would not buy it to “match Muse.” Muse does not use it.
Mac mini M5 Pro 24 GB on Amazon
ASIN B0HGGHNQY6. Amazon Associates link. The live price is on that page, not locked in here.
Mac Studio M5 Max 36 GB on Amazon
ASIN B0HGKSQMX6. 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’s announcement says Muse runs on a Muse Secure VM in the cloud, powered by Muse Spark. The open-weight model you can run locally is Muse Glimmer, a separate 30B release. Adobe Firefly and Black Forest Labs FLUX are not Muse.
The dots feature page does not describe a free tier. Access listed there is Pro, Business Premium, and Enterprise, and Enterprise requires an admin. The name is lowercase dots. Operator is not the current product on that page.
No. Grok Bot, two words, is xAI’s always-on teammate with a cloud computer. The desktop and iOS apps are how you message it. The introducing post lists SuperGrok tiers and paid Cursor plans, with enterprise on a waitlist. It does not say you can download Grok weights.
A 7B Q4 (about 6.2 GB) is a light helper, not a replacement. A 27–32B Q4 (about 18–21 GB) on 24–36 GB of memory is the first tier that can attempt the same chores with an open model. 128 GB is for a large MoE, not for “three cloud agents.”
A correction or a hardware question goes to the contact page.