Make AI your own. Local models + massive storage. Expand when ready. No cloud bills, no token fees. $10 to lock $400+ off ↓
Hong Kong
Joined October 2026
- Tweets94
- Following21
- Followers235
- Likes37
A closer look at how XAIStack’s compute module connects to the main unit, with the system powered off.
What should we show next: GPU installation or the drive bays?
819
Why We Chose This CPU
Two questions keep coming up: “Isn’t that a laptop chip?” and “Can I swap it later?”
We chose the Intel Core Ultra 5 deliberately. In XAIStack the CPU is the hub -- it connects the GPU, system memory, and fast SSDs at full speed so big models can run the way we designed.
Three things mattered most:
1. Lanes
28 PCIe lanes. Full PCIe 5.0 ×8 to the GPU (shared with nothing), four M.2 SSDs that don’t share lanes, dual 10GbE + six SATA bays. More than the 16-lane chips in many unified-memory boxes.
2. Memory you can grow
Two SO-DIMM slots. X1 up to 96GB, X1 Pro up to 128GB. Start small, add more later.
3. Power for 24/7
28W base power. Main unit idles ~35W. When a model runs, the heavy power draw stays on the GPU.
Result from our tests: same CPU + 64GB RAM, just swap the GPU -- 50+ tok/s on a 125B model (RTX 5060 Ti), 30+ tok/s on a 763B model (RTX PRO 5000).
CPU is soldered in this generation. When we release a newer host, you swap that one module. GPU, RAM, SSDs, drives and other modules carry over.
👉 Reserve for $10: hello.xaistack.hk/reserve-pa…
Discord: discord.com/invite/369BwtJyA…
2
2
1,459
Who builds XAIStack, and how will support work after delivery?
XAIStack is developed and manufactured by Centerm, founded in 2002. Centerm ships to 120+ countries and ranked first in global thin client shipments in 2023–2025, according to IDC data cited by Centerm.
1
4
1,914
Our manufacturer’s warranty covers the system for 1 year and specified core components supplied by us for 3 years, including CPU, memory, SSD, power adapter and power supply.
For covered faults, replacement parts and shipping to you are free. Accidental damage is excluded.
1
2
73
An early look at XAIStack’s performance and IOMMU groups from @jimsgarage_. Thanks for putting it to the test!
1
800
Some exciting news before launch. 💚
XAIStack has been selected as a @Kickstarter Project We Love!
We're grateful for the recognition and even more excited to show you what we've been building.
Launching November 12.
#ProjectWeLove
1
563
A little behind-the-scenes look from #Singapore 🇸🇬
We're getting ready for #EDUtechAsia, Nov 4–5!
Come meet the XAIStack team at Booth D02 (Centerm) and experience local AI in person.
See you there!
#EDUtechAsia #LocalAI #XAIStack
414
Same host. Different builds.
Start with the main unit. Add a compute module, a storage module, or both.
You don't need to build the biggest setup on day one.
Which configuration makes the most sense to you?
1
1
322
Cloud cameras upload your footage.
Then charge you monthly to look at it.
Run detection locally.
Keep the video on your own storage.
No cloud. No subscription.
ALT XAIStack modular AI workstation on a desk, connected to a large monitor showing multiple security camera feeds with AI detection boxes around people and vehicles.
1
248
Building a private, on-premises data center has long been a challenge, especially when you want AI to search, organize, and answer questions about locally stored data.
A common approach is to store files on a NAS and run AI models on a separate GPU-equipped PC. This adds work: managing access between devices, configuring permissions, and maintaining both systems.
XAIStack brings NAS storage and AI computing together in one desktop device, with modular GPU and storage expansion to simplify deployment. Most importantly, no need to uploading your documents to the public cloud.
#LocalAI #PrivateAI #NAS #OnPremises #DataPrivacy #XAIStack
3
203
We recently ran a 125B model locally on XAIStack and asked it to create “a magnificent pagoda.” The whole task took 3 minutes. It handles longer tasks just as well as the original full-precision model.
#XAIStack #LocalAI #LLM #AIInference #AIWorkstation
2
207
The token consumption of AI agents is on a completely different level compared to standard chat interactions. Combining cloud-based and local models offers a much superior solution.🤔
token cost could go down 1000x and our enterprises would just stuff in 1000x more tokens into the models
the need for intelligence is unbounded in comparison to the few things we process today
151
Local AI models are thriving! Another great candidate for running locally—can’t wait to test out its performance.😀
Here is an open-weight coding model you can run locally on your computer.
It's a 40B parameter model with a mixture-of-experts architecture. It activates 2.5B parameters per generated token. 512K context window.
This thing runs on regular computers! 125 tokens/sec on an M4 MacBook Pro.
Good alternative to closed models, but make sure you check their usage guidance to understand the trade-offs.
135
