System76 Thelio Mira Custom Review: A Quiet, Powerful Linux Workstation

System76's 2026 Thelio Mira Custom is a Linux desktop starting at $1,699, built on AMD's Ryzen 9000 platform with an X870 motherboard and topped with an Nvidia GeForce RTX 5070. ZDNET reviewer Steven Vaughan-Nichols tested a 16-core Ryzen 9 9950X configuration and called it the finest ready-to-run Linux workstation he has ever used. Liquid cooling makes it near-silent even under heavy load, and System76 claims a 13-degree temperature drop and 19% performance gain over the previous air-cooled generation.

What happened

Spec	Detail
Price (base)	$1,699
CPU range	Ryzen 5 9600X (6-core) to Ryzen 9 9950X3D (16-core)
Reviewed CPU	Ryzen 9 9950X, 4.3 GHz base, 5.7 GHz boost
GPU	Nvidia GeForce RTX 5070, 12 GB VRAM
Motherboard	X870, PCIe 5.0
PCIe 5.0 throughput	64 GB/s (128 GB/s bidirectional)
NVMe Gen5 sequential speed	Up to 15 GB/s
Cooling	Liquid cooled; 19% perf gain, 13°C cooler than prior air design
OS	Pop!_OS (System76’s own Linux distro)

System76 refreshed its Thelio Mira desktop in 2026 around AMD’s Ryzen 9000 “Granite Ridge” CPU family. The X870 motherboard supports PCIe 5.0 lanes for both the GPU and SSD, giving 64 GB/s of bandwidth per direction. The Gen5 NVMe slot adds sequential read speeds up to 15 GB/s, which is roughly twice what most current SSDs manage.

The reviewer’s unit came with the top-of-the-range 16-core Ryzen 9 9950X. Under everyday tasks it ran at 4.3 GHz, jumping to the advertised 5.7 GHz boost ceiling when pushed hard. The RTX 5070 carries 12 GB of VRAM, which covers 1080p and 1440p gaming but falls short for serious 4K competitive play.

Why the liquid cooling matters

The big news in this generation is the move from air to liquid cooling. System76 says the change brings a 19% performance improvement and a 13-degree Celsius temperature reduction. In practice, the reviewer reports the machine was effectively silent even while running demanding workloads, to the point of not being sure the PC was powered on at all.

That silence is not just a comfort feature. Sustained cooling directly supports long AI inference and fine-tuning runs. A machine that throttles after 20 minutes of GPU stress is nearly useless for batch generation or multi-hour training jobs. The Thelio Mira is designed to hold its boost clocks throughout those sessions.

Is the RTX 5070 good enough for local AI?

For running large language models (LLMs) locally on the GPU, 12 GB of VRAM places this card in a specific performance band:

• Comfortable: 7-8B parameter models such as Llama 3.x/3.1 8B, Qwen2.5-7B/8B, and Mistral Small 3 7B.
• Tight but workable: 12-16B models like Qwen3/Qwen2.5 14B, Llama 3.x/3.1 15B, and Gemma 3 in the 12-14B range.
• CPU offload: The Ryzen 9 9950X can handle AI tooling, fine-tuning, and mixed workloads on the CPU side when the GPU alone is not enough.

Pop!_OS also allows GPU switching on the fly via a graphical interface. You can run display output from the Ryzen’s integrated graphics while pointing specific apps at the discrete RTX 5070, which is a practical way to dedicate all 12 GB of VRAM to inference rather than desktop rendering.

If you are researching which local models fit a 12 GB card, our roundup of the best local LLMs for 24 GB GPUs also covers the smaller memory tiers and gives useful comparisons across Qwen, Gemma, and Mistral families.

Pop!_OS and the driver situation

One of the persistent frustrations with Linux desktops is Nvidia’s proprietary CUDA drivers. System76 ships Pop!_OS with those drivers already installed and tested. Additional CUDA tools can be added through Pop!_Shop or a single terminal command, without dependency conflicts. That is a meaningful quality-of-life difference for anyone who has spent an afternoon fighting Linux driver installs.

Our take

The Thelio Mira Custom is genuinely compelling if you want a pre-built Linux machine you can trust for AI workloads. The spec list is honest: the RTX 5070 is an entry to mid-range local inference card, not a powerhouse. If you are running 30B+ models or serious 4K gaming, you will want to configure a higher-tier GPU.

What System76 gets right is the whole-system story. Liquid cooling, a strong CPU for CPU-side AI work, a Linux OS that actually includes working Nvidia drivers, and a case designed for repair and expansion. At $1,699 for the base config, it is not cheap, but compared to the hours you would spend building and debugging an equivalent custom rig, the premium is defensible.

For businesses using AI tools locally, such as on-premises inference to keep data private, or teams doing prompt engineering and fine-tuning, this kind of machine pairs well with a structured AI integration strategy rather than a one-off hardware purchase. Hardware is only as useful as the workflows built around it.

What to do about it

1. Decide on your primary workload: gaming, AI inference, or both. That determines whether 12 GB VRAM is enough or whether you need to configure a higher-spec GPU at order time.
2. If AI inference is the goal, check which model sizes you plan to run before ordering. 7-8B models fit fine; 14B+ models will use CPU offloading.
3. Consider the upgrade path. The X870 board and PCIe 5.0 lanes give room to swap in a larger GPU later without replacing the whole system.
4. Budget for peripherals. System76 does not bundle a display, keyboard, or mouse, so factor that into the total cost.

If you need GPU-accelerated AI tooling that just works on Linux, the Thelio Mira Custom is the most turnkey option currently on the market.

Source: ZDNET · AI