Ghost has launched Core, a personal AI computer priced at $3,499 that runs its models entirely on the device, with no subscription. Batch 1 is already sold out, and the product page now asks visitors to register interest for the next one. The pitch is a single line: "Intelligence you own."
The idea is bigger than a local chatbot. Ghost describes a box that runs 24/7, connects to your email, calendar, files, browser history, wearables and smart-home gear, builds context on your life over time, and acts "without waiting for a prompt." That is the same always-on assistant direction as OpenAI's Dots, except the compute and the data stay in your house. This post covers what Ghost has said, what it has left blank, and how Core should be judged against the cheaper route of building your own local AI machine.
TL;DR: the questions people are asking
| Question | Short answer |
|---|---|
| What is it? | A dedicated personal AI computer that runs open-weight models locally and continuously |
| Price? | $3,499, no subscription |
| Can I buy one? | Batch 1 is sold out; you can register interest |
| Which models? | Qwen 3.8-Next, Qwen 3.8-27B, Gemma 4-31B, Muse-Glimmer-30B |
| Specs? | Not published on the launch page; GPU, CPU, memory and storage are listed without values |
| Does data leave the device? | Ghost says no; this is unverified |
| Returns? | 30-day returns |
| What does it connect to? | Email, calendar, finances, files, recordings, browser, lights, fridge, speaker, cameras, health records, Whoop, Oura, Eight Sleep |
| Is it proactive? | Yes, by design; it surfaces suggestions without being asked |
| Should I wait for reviews? | Yes, for speed, noise, power draw and real privacy behavior |
What Ghost actually announced
Everything below is from Ghost's launch page and video, not from hands-on testing. We have not used the device.
Local-only operation. Ghost says "everything you do with Core is run locally using models on-device," that "no data ever leaves your home," and that your AI "is owned entirely by you." There is no subscription, which implies no recurring inference bill.
Continuous context. Core "runs 24/7, continuously taking in the context of your life." The page lists inputs by category: digital sources such as email, calendar, finances, files, recordings and browser; home devices such as lights, fridge, speaker and cameras; and health data from health records, Whoop, Oura and Eight Sleep. Connectivity is over Wi-Fi and Bluetooth.
Proactive behavior. The page shows example nudges at specific times: a higher recovery score than usual at 6:48 AM, a reminder about a package arriving that day, a note about running low on groceries and rain changing the commute at 2:15 PM, and a remark that you have been eating tacos for four days straight. Another set of examples suggests taking an apple because you said you wanted to eat more fruit, drinking water after two coffees, and not forgetting your keys. These are marketing illustrations, not documented features, so treat them as intent.
Models. Four are named: Qwen 3.8-Next, Qwen 3.8-27B, Gemma 4-31B and Muse-Glimmer-30B. We have covered the open-weight side of that list, including the Qwen 3.8 27B release and Meta's Muse-Glimmer 30B.
Terms. $3,499, 30-day returns, and batch-based availability.
What Ghost has not said
The launch page shows the headings GPU, CPU, memory and storage without any values under them. For local AI, those are the only specs that matter, so the gaps are significant.
| Missing detail | Why it matters |
|---|---|
| Memory capacity | Sets which model sizes fit and how much context you can keep loaded |
| Memory bandwidth | Largely determines tokens per second when generating text |
| Chip and power draw | Decides noise, heat and the electric bill for a machine that never sleeps |
| Measured speed | No tokens per second figure for any of the four models |
| Ship date and shipping regions | Batch 1 is sold out, with no timeline for the next |
| Software stack | Whether you can run your own models, read the code, or export your data |
| Privacy architecture | How context is stored, encrypted, and deleted |
| Action permissions | What the box may do on its own versus ask first |
Until these appear, a fair reading is that Core is a product announcement and a demand test. Batch 1 selling out tells you there is interest, not that the device performs.
How does the model lineup run at home?
The four models are in the 27B to 31B class, plus a "Next" variant. Open models of that size are practical on a well-specified local machine, usually with quantization to cut memory use. A 30B-class model at 4-bit precision needs on the order of 15 to 20 GB of memory for weights alone, plus room for the context cache. That is a rule of thumb, not a Ghost figure.
What this means in practice:
- Quality is capped by the open-weight frontier. Local models are good and improving, but they are not the same as the largest hosted systems. Ghost is trading peak capability for ownership.
- A single box serving many tasks shares hardware. If Core is ingesting your email and watching your sensors continuously while you also ask it hard questions, those jobs compete for memory and compute.
- Continuous context is mostly retrieval, not training. Models do not learn your life in their weights. They read from a store of your data. How good that store and its search are matters more than the model name.
For readers weighing hosted against local, our closed-source versus local open-source comparison lays out where each side wins.
Core versus building your own
At $3,499, Core lands in the territory where you could assemble or buy a capable local AI machine yourself. The comparison below uses only categories, since Ghost has not given specs.
| Factor | Ghost Core | DIY or off-the-shelf local box |
|---|---|---|
| Setup effort | Minimal if it works as advertised | Hours to days |
| Known specs | Not yet | Yes |
| Model choice | Four listed models; flexibility unknown | Any open-weight model |
| Always-on assistant layer | Built in | You assemble it from open tools |
| Upgrades | Unknown | Often possible |
| Privacy proof | Vendor claim | You control the stack and can inspect it |
| Support | Vendor | Community |
Our guides cover the DIY side in depth: NVIDIA DGX Spark for local LLMs, a MacBook versus dedicated GPU comparison, and how to build a personal AI system with a local workflow. The wider hardware market is also moving, as in the DGX Spark $4,999 memory-cut story, so memory pricing could shift what $3,499 buys.
If you like the always-on assistant idea but not the closed box, open projects already cover much of it. Start with what OpenClaw is and how it works and the MemPalace local memory project.
The privacy question deserves scrutiny
"No data ever leaves your home" is the strongest claim on the page, and also the one hardest to verify. Several questions follow from the connector list alone, which includes finances, health records, home cameras and screen history.
What is stored, and for how long? Continuous capture builds a detailed record. Retention, deletion and export controls decide whether that record is an asset or a liability.
Is the device encrypted at rest? A box full of your personal context is a high-value target if someone steals it or gains access to your network.
Do updates or telemetry phone home? A local model can still send diagnostics. A clear policy and an independent network trace would settle this.
What happens if Ghost shuts down? Hardware that depends on vendor software for updates can become a brick. Open software or exportable data mitigates that risk.
Who can the assistant act as? A system that reads email and calendars and can take actions needs strict permission boundaries. We have seen what happens when assistants act unprompted, as in our coverage of ChatGPT Dots sending an unapproved email.
None of this is a reason to dismiss the product. It is the checklist to apply when the first reviews and teardowns arrive.
Is a proactive AI what you actually want?
The demo nudges are mild: bring an apple, drink water, check the weather. The harder question is where proactive help stops being useful and starts being intrusive. A message that notes you are most stressed on Sunday nights, or that you have eaten tacos for four days, treats your behavior as data to comment on. Some people will find that valuable. Others will find it unsettling, even if everything stays on a local drive.
The practical test is control. Can you set the topics Core may raise? Can you mute categories, delete what it has inferred, and see why it said something? A good proactive system explains itself and is easy to quiet. The launch material does not show those controls, so reserve judgment.
What this means for what you build
If you build with AI, Core is a signal more than a purchase. Open-weight models in the 27B to 31B range are now good enough that a company is willing to sell a consumer device around them with no subscription. That shifts the baseline for what you can ship privately. Apps that assume a cloud call for every inference may soon compete with products that run it on the user's own hardware.
If you are deciding whether to buy, the sensible path is to register interest, wait for specs and independent reviews, and compare price per gigabyte of fast memory against the DIY options above. If you need a local assistant today, you can build one with open tools and the same model families Ghost lists.
Related reading on explainx.ai
- What is TPS? Tokens per second explained
- Qwen 3.8 27B open-weight model compared with Claude Opus
- Meta Muse-Glimmer: open-weight 30B agentic model
- NVIDIA DGX Spark: best local LLM setup
- MacBook vs dedicated GPU for local LLMs
- Build a personal AI system with a local workflow
- Closed-source AI vs local open-source alternatives
- OpenAI Dots and always-on agents
Details are based on Ghost's launch page and video as of October 5, 2026. Specs, availability and pricing may change; we have not tested the device. Check Ghost's official channels for current information.