Motivation

Cloud language models are suitable for a wide range of tasks: writing code, analyzing data, and answering technical questions. The author identifies four problems with cloud services: model hallucinations, data privacy, the use of human-created content to train models without the authors' consent, and token costs.

Running models locally solves some of these problems. Buying hardware upfront eliminates ongoing token costs, but does not eliminate them entirely. Some tasks still require state-of-the-art cloud models. According to the author's estimate, more than 90% of the necessary tasks can be performed on your own hardware without a subscription.

The main advantage of a local workstation is the ability to connect always-on computing to other systems without the risk of overspending on tokens. The data stays within the local network, and the service works without an internet connection.

Case

The main reason for the rebuild was the release of the updated version of the Silverstone Temjin TJ07 case. The author calls the previous model his favorite case and used it for complex water-cooled builds.

Key features of the case:

• an aluminum body made from a single solid sheet, except for the flat rear panel;
• mounts for multiple graphics cards, designed for multi-GPU systems;
• space for three radiators: in place of the front 5.25-inch bays, as well as at the top and bottom;
• support for multiple power supplies for high-power systems;
• wheels at the base.

The author notes that gaming system companies need fewer graphics cards and bulky tower cases, whereas AI workstations do need them.

Components

The processor and motherboard belong to the Intel Sapphire Rapids platform. The author selected it from hardware that had not been used in other lab projects. The choice of processor was driven not by clock speed but by support for a large amount of memory and a large number of PCIe Gen 5 lanes. The author plans to use these lanes when running a large number of containers and installing multiple graphics cards. On the motherboard, all PCIe slots operate at full x16 width.

The graphics cards were taken from another lab system. One of them previously received a homemade solution for the 12V-2x6 power connector, using an XT120 adapter. The author acknowledges that the solution was inexpensive, but it works.

The memory comes from a kit that had been set aside on the top shelf of the storage room. The author nevertheless used the boxes labeled "Do not take without permission from the business team." RDIMM memory is currently expensive.

Processor Installation

The author installed an LGA 4677 platform processor for the first time. According to the author's description, thermal paste is applied to the processor, after which the processor is secured to the cooler. Tightening is performed in stages, in a crosswise pattern. For some server platforms, tightening fully in a single pass is permitted, but this method is not suitable for this platform. The tightening torque is 8 inch-pounds (approximately 0.9 N·m).

Motherboard Installation

During installation, compatibility problems came up. The protective plate on the back of the motherboard and the rubber cable grommets were thicker than required. The author flipped the grommets: they are thinner on one side, and this resolved the conflict.

The standoffs for the motherboard came without instructions. The author could not find a mounting template, so they positioned the standoffs manually.

The case has a cable routing hole that is wide enough for a screwdriver. The intake fan is installed at the front. The airflow is arranged so that air is exhausted through the top of the case.

Cooling

The author chose air cooling for the CPU because few water blocks are available for this platform. The processor is not overclocked, and the temperatures achieved with the air cooler are sufficient.

The water cooling loop for the graphics cards includes:

• a 420 mm radiator found in storage; it is the only such radiator on the premises;
• a pump and reservoir mounted on the front panel of the case; the panel has mounting holes with a 120 mm spacing;
• ZMT soft tubing with low permeability to evaporation;
• Bitspower True Silver silver fittings, about 20 years old; silver has biocidal properties and helps keep the loop clean.

The author notes that high-quality water cooling components can be used for many years if one does not follow constantly changing trends.

Power Supply

The Silverstone Zeus 1650R Titanium power supply connects to two 12V-2x6 connectors. For a system with two RTX 5090 graphics cards, the author chose a 1650 W unit. The case has room for a second power supply if the power proves insufficient.

The specifications of the 3000 W Silverstone power supply indicate a 220–240 V input. The manufacturer does not mention operation on 120 V mains. At 120 V, the unit does not deliver its full power, because the current would exceed the rating of the circuit breaker. The author assumes the unit may fail to start at 120 V.

The motherboard requires:

• a 24-pin power connector;
• up to two 8-pin EPS connectors;
• one 8-pin and one 6-pin PCIe connector for supplemental slot power.

Each PCIe slot delivers up to 75 W. With seven slots, the total power can reach 525 W. This load exceeds the capacity of the two 12-volt wires in the 24-pin connector, so additional connectors are mandatory when installing a large number of graphics cards.

The 12V-2x6 connector provides data exchange between the power supply and the graphics card. The unit determines the graphics card's power consumption and confirms that it can be supplied. A homemade cable does not contain these communication lines. To simulate the signal, the author shorted the signal wires with jumpers. As a result, the graphics card receives permission to draw 600 W.

Assembly Error

During assembly, one of the parts was installed incorrectly: the rear fan mounting opening protruded inward into the case, into the area of the I/O connectors. Fixing it required removing the motherboard, installing an additional fan, and a fan splitter.

Software Setup

The software stack consists of three components:

1. LM Studio: an environment for running and hosting local models. The model is accessible on the local network and can be hosted on a separate machine.
2. Qwen 3.8: a model running on both graphics cards. The context is about 108 thousand tokens. The model occupies 25–29 GB of VRAM and fits on a single graphics card, but LM Studio distributes it across two. According to the author, the default parameters turned out to be optimal.
3. Hermes: an agent shell that wraps the model, sets the system prompt, and connects tools and integrations.

Hermes is connected to Discord; Telegram and other messengers are also supported. The bot was named Sventana. When running, the agent performed tasks in the background: it checked the news and compiled a digest. The author notes that such tasks can be scheduled in advance so that the morning digest is ready when the user arrives.

Processing speed is low. Responses are generated slowly, but background tasks such as writing code, testing, and fixing bugs can run without user involvement, including overnight.

Additional Hardware Solutions

To work with large models, the author considers two options:

• One fast GPU with additional cards. Instead of two RTX 5090s, one card can be used for fast work and several cards with increased memory capacity for tasks where speed is not critical. The author mentions graphics cards whose memory has been increased after modification; one of them, according to him, grew from 8 to 64 GB.
• DGX Spark. The laboratory has received these systems. According to the author, they rely on memory capacity rather than processor performance. One of the systems runs the DeepSeek 4.1 Flash model with roughly 410 GB of parameters.

Summary

According to the author, the local station allows:

• performing most work tasks without a subscription or token payments;
• storing data on the local network;
• operating without an internet connection;
• running background agent tasks without constant user involvement.

Limitations noted by the author:

• cloud-based frontier models are still necessary for some tasks;
• local models are slower than cloud models;
• the cost of components, especially memory and power supplies, is high;
• assembly requires compatibility checks and careful planning of space for cables and connectors.

The author links the next stage to integrating the station with Home Assistant. He plans to use object recognition at the entrance to the house with notifications. The author notes that motion sensors fail at this task in heavy rain: during rain, the light turned on automatically. A solution may be a rule that takes weather data into account.