Oki Home: A $1,799 "Memory Computer" With a Local Qwen 3.8 27B and a Swappable Memchip

The idea of a personal AI that knows your life and never phones home has been a staple of demos for years. On October 6, 2026, Oki, a Y Combinator Winter 2025 company, put a price and a date on it: Oki Home, a $1,799 box that stores your photos, messages and files on a swappable Memchip, indexes them on one timeline and answers questions about them with a local Qwen 3.8 27B model. Reservations cost $59, and shipping is planned for mid-December 2026.

This post lays out the specs, explains what the product does, compares it with building a similar machine yourself, and lists what is unproven. Everything here comes from the founder's launch thread, the product page and early press coverage. We have not tested the device.

The announcement

Founder Luofei Chen introduced Oki Home in a thread, with a short demo video, saying the team built it "because we felt the old promise of the personal computer slipping in the new age of nearly boundless intelligence." The pitch is a deliberate inversion: instead of products that "optimize for maximal data collection and analysis while offering limited compute," Oki offers a machine with plenty of local compute that keeps data at home.

TL;DR: the questions people are asking

Specs, in one place

The Memchip is the distinctive piece: a credit-card-sized M.2 module, about 86 x 54 x 6 mm with a titanium rim and aluminum heatsink, that holds the user's data. The framing from one reply captured the logic: "the model can be replaced, the years of context it learns about you can't," to which the founder answered that it is all stored on the Memchip. If that holds, upgrading the machine or the model would not mean rebuilding your personal index.

What the software is meant to do

Three features define the product.

1. A timeline of your digital life. "Every photo, message, and file lines up on one timeline, automatically organized." Data arrives through apps and integrations. The product page lists connectors for Google Calendar, Gmail, Google Drive, Apple Messages, Apple Photos, Apple Health, Outlook, Notion, Dropbox and Zoom, among others.

2. Search in plain language. "Ask in your own words. Oki finds it." You can ask for a moment or document by who, where or when, and the system uses the indexed archive.

3. Oki Desk, an agent. Oki Desk can browse websites, compare options and work through multi-step tasks using your personal context. Coverage notes this is where local and external services meet: browsing the web requires leaving the box.

There is also an open side: a Linux environment for building custom apps and running your own AI models, so the hardware can do other work. Updates, including system and AI improvements, require user approval.

The performance claim, and what 106 tokens per second means

Oki says its optimized Qwen 3.8 27B runs at 106 tokens per second output and 1,176 tokens per second prefill on this hardware. Context helps. A 27B model at 4-bit quantization takes roughly 14 to 16 GB of weights, which is tight but possible on a 16 GB card with a modest context window. Other builders have reported similar throughput using speculative decoding and careful quantization, and a cloud vendor demonstrated far higher speeds on specialized hardware, as in our Cerebras Qwen 3.8 27B coverage.

Three cautions apply. First, this is the founder's number and no independent benchmark exists yet. Second, throughput varies with quantization, context length and whether the machine is also indexing photos or running an agent. Third, a fast model does not guarantee good retrieval: the quality of "find the moment I mean" depends on indexing, embeddings and ranking more than on the language model's speed. For the model itself, see our post on Qwen 3.8 27B and how it compares with Claude Opus.

How it compares with building your own

A fair question is whether you could assemble this for less. The parts are ordinary consumer hardware.

The honest value proposition is not the hardware, which you can buy elsewhere. It is the integrated product: connectors, a unified timeline, a working search experience, an agent and updates, packaged so that a non-technical person can use it. Whether the software delivers is the open question.

Privacy: what the claims cover and what they do not

Oki's central promise is that "your questions and your data never leave your home." There are good reasons to find that plausible: a local model, local storage, a device you physically own. There are also places where you should read closely.

• Connectors pull data in, and some push data out. To index Gmail, Drive or Apple Messages, the device must authenticate to those services. Check what each connection stores, for how long and with what scope.
• Oki Desk goes to the web. Browsing and comparing options sends requests to external sites, and any personal context included in a task could be exposed to them.
• Remote access. Reaching the box from a phone or laptop away from home needs a secure path. Ask how it is built and who operates any relay.
• Updates. Over-the-air updates require approval, which is good, but you should know what telemetry, if any, the device sends.
• Physical security. A Memchip that holds years of your photos and messages is a very valuable object. Ask about encryption at rest, key management and what happens if the chip is lost or the device is stolen.
• Backups. One box with all your memories is also one point of failure. Plan a backup.

The same lessons apply as for any agent that holds personal and work data, as we discussed in our posts on agent permissions and shared connectors and on the Underdog private personal AI launch. Treat a "private" device as a claim to verify, ideally with an independent security review.

Who it is for, and who should wait

Reasonable early adopters: people who already keep large photo and document archives, want search over them and distrust cloud services; tinkerers who like the Linux environment and local models; small teams that want an on-premises assistant.

People who should wait: anyone who needs the product to work reliably on day one, since this is a first batch from an early-stage company; anyone who would connect highly sensitive accounts before a security audit; buyers uncomfortable with a refundable but non-trivial pre-order.

A sensible approach is to reserve if curious, since the $59 is reported refundable, and to hold off on a full commitment until reviewers have tested search quality, speed and privacy behavior in the real world.

Questions to ask Oki

1. How is the 106 tokens per second measured, at what quantization and context length?
2. What is the indexing pipeline, and how does it handle photos, messages and documents at scale?
3. How accurate is search on an ambiguous query, and is there a way to see why a result matched?
4. What data does the device send off the box, by feature?
5. How is the Memchip encrypted, and can I recover data if the device fails?
6. What happens to my data and the device if the company shuts down?
7. How many units are in the first batch, and what is the refund process?
8. Can I run other models, and is the Linux environment fully open?

What this means for what you build or pay

For consumers, Oki Home is a bet that people will pay a premium to own their AI and their memory. For builders, it shows a product category forming around "personal memory plus local model," a space that also includes work-memory tools such as Screenpipe and agent computers such as Pamir's Lapis One. If you build in this space, notice the pattern that Oki emphasizes: hardware you can touch, a storage module that carries identity, and an explicit privacy promise. If you are shopping, wait for independent reviews before paying more than the reservation.

Related reading

• Qwen 3.8 27B vs Claude Opus: the local model
• Cerebras Qwen 3.8 27B at 1,500 tokens per second
• Strata: a 125B model on a 12 GB GPU
• MacBook vs dedicated GPU for local LLMs
• Underdog: a private personal AI launch
• Pamir Lapis One: an agent computer
• Screenpipe: local work memory for agents
• What is MEMORY.md? Agent persistence explained

Primary: Luofei Chen's launch thread on X (October 6, 2026) and demo video · okihome.ai · RuntimeWire coverage of the launch

Details are accurate as of October 7, 2026 and come from the founder's thread, the product page and press coverage. We have not tested the device. Specs, pricing, shipping and privacy behavior may change, and performance numbers are the vendor's claims.