eRacks Systems

We spent some time with MiMo Code, an open-source terminal-native (runs from the command line, not a separate app window) AI coding assistant that reads and writes code, runs commands, and manages Git while keeping memory of your project across sessions.

It held up well enough that we're comfortable installing it for customers who want a coding assistant running locally on their own hardware instead of piping code through someone else's cloud.

This is the kind of thing we do routinely: try the tool ourselves first, then set it up on the servers we build so you're not stuck configuring it alone.

https://buff.ly/3VlO8NH

There's a project called colibri running a 744 billion parameter (the numbers a model learns from training) model called GLM-5.2 on a 25GB machine with no GPU (graphics processing unit, usually the expensive part of an AI rig). It works by keeping the core of the model in RAM and streaming the rest from disk as needed.

It's slow. Disk speed is the bottleneck. But it's a real proof that you don't always need a rack full of GPUs to run a huge model, you need the right architecture and fast enough storage.

This is the kind of thing we pay attention to when we spec on-premise AI boxes. Sometimes the answer is more RAM and better NVMe (a fast type of solid state storage) drives, not another GPU.

https://buff.ly/Fr47CFq

Saw this one on TechPowerUp: a single-slot GeForce RTX 3060 (a mid-range GPU, graphics processing unit) with no separate power connector, drawing all its power straight from the PCIe (the slot itself), listed for $495 on Newegg.

That form factor is genuinely useful for dense rackmount builds where slot space and cabling are tight. We deal with that tradeoff constantly: compact cards are great until you look at the price and availability (p&a) and realize the math doesn't always work out compared to a standard dual-slot card.

Still, it's worth watching. Single-slot GPU options are rare, and for certain server chassis, that rarity alone might justify a premium if the price settles down.

https://www.techpowerup.com/352447/single-slot-geforce-rtx-3060-with-no-power-connector-listed-on-newegg-for-usd-495

An XDA writer kept his Synology NAS and replaced nearly every Synology app with open ones: Immich for photos, Navidrome for music, Plex and Jellyfin for video, Paperless for documents, all in containers. We have been saying this for years: once the apps are open, the box is just storage, and the storage should be open too. ZFS, standard drives, your choice of Linux. https://blog.eracks.com/2026/09/open-nas-apps-not-the-box/

ServeTheHome spotted AMD's Ryzen AI Halo, an AI developer PC (a small workstation built for running AI models locally) at Computex, running on the Ryzen AI Max+ 395 chip that's actually been shipping for almost a year already.

Look for this one. It's a sign that local, on-premise AI hardware for developers is becoming a normal product category, not a science project.

We've been building private, on-prem AI servers and workstations at eRacks for a while now. This is just more validation that the market's catching up to where the demand already is.

https://www.servethehome.com/we-found-the-amd-ryzen-ai-halo-ai-developer-pc/

Saw this thread going around: 11 tools and habits for not wasting your Claude context window (the chunk of text an AI model can 'see' at once). The MCP servers (Model Context Protocol, a way to feed live tools and data into Claude) are the useful part: serena reads code by symbol instead of whole files, context7 pulls current docs so Claude stops guessing at APIs, claude-context does semantic search so it only pulls in what matters.

We install this stuff. If you're running Claude Code on your own hardware, this is exactly the kind of setup work we do, get the MCP servers wired in, get the context window actually working for you instead of getting chewed up on stuff it doesn't need.

Worth a read if you're using Claude Code day to day.

https://www.threads.com//post/DZP_iLWDqY0?xmt=AQG05nY9P77K22cI7j69ylR7h0DB6nTLFX8kXDXr81N9AQ2KKe73weBg4cc8B-3G5B4Fjhg&slof=1

Good article on why Google's new 'Googlebook' platform (a Chrome OS/Android hybrid with heavy AI features) is missing the point. The piece argues people don't want AI bolted onto their OS as the headline feature, they want tools that just work.

We've been saying some version of this for 25 years. A computer's job is to run your software reliably and get out of your way. That's true whether it's a rackmount server or the laptop on your desk.

If you want a real Linux desktop or laptop, built to spec, no gimmicks, no mystery AI layer trying to guess what you meant, that's just a normal order for us.

https://www.xda-developers.com/magic-pointer-wont-save-googlebook-because-google-doesnt-understand-pc-users-want/

Clear Linux (Intel's performance-tuned distro) is on its way out, and folks are asking what replaces it. CachyOS is shaping up to be a good answer: Arch-based (built on Arch Linux, a rolling-release distro), tuned for speed, and flexible enough to run whatever desktop environment (the GUI layer you interact with) you like.

We've been building custom Linux servers and workstations since 1999, so distro transitions like this are just part of the job. If you want CachyOS on your next rackmount server, workstation, or AI box, we'll install it.

Worth a look if you're weighing options for your next build.

https://search.app/5HDqs

Steam on Linux just hit an all-time high, 3.2% of all Steam users according to the November survey. A decade ago it was under 1%. Proton (Valve's compatibility layer that lets Windows games run on Linux) and the Steam Deck have done a lot of the heavy lifting here.

We've been building Linux desktops and workstations for 25 years, and Steam support has been part of that the whole time. When you order one of our Linux desktops, Steam is just one of the options in the dropdown. Pick it, we install it, done.

No separate setup, no forum threads, no fiddling with Proton compatibility settings after the fact. It's already there when the box shows up.

https://buff.ly/lINuRNQ

Saw this piece on "Claude Code Ultrawork," where people are running AI coding agents (software that writes and tests code on its own) autonomously for hours, spinning up parallel agents to build whole apps without stopping.

We will keep watching where this goes. Whether it's an official feature from Anthropic or a community workaround, the pattern is clear: people want AI agents that run longer, do more, and don't need constant babysitting.

That's exactly the kind of workload we think about when we spec out on-premise AI servers. Long-running, parallel, resource-hungry jobs need real hardware underneath them, not just a laptop fan spinning up.

https://buff.ly/bQuXZUk