Dedicated local AI machines: a comparison of small PCs built to run models at home

The landscape of machines built specifically for local AI has gotten noticeably denser in 2026. Here is a panorama of the main candidates, with for each the maximum supported memory, an indicative price range as of today and what you can or cannot upgrade yourself.

Overview

Apple Mac Studio (M5 Max / M5 Ultra)

The Mac Studio remains, on memory, the unquestioned reference in this niche. Announced on August 25, 2026, the M5 Max version starts at $3,300 with 36 GB of unified memory and climbs to 128 GB, with a memory bandwidth of 614 GB/s. The M5 Ultra starts at $7,200 with 96 GB, reaches $13,600 in its 256 GB configuration and will offer a 512 GB option (price not yet published) from the end of October, with a 1.2 TB/s bandwidth. Apple itself states that the M5 Ultra runs "frontier-class" open models on device, and MLX is the ecosystem that makes the best use of this architecture, with quantized models running at very usable speeds.

• Maximum memory: 512 GB (unified memory, M5 Ultra — 512 GB version shipping late October 2026).
• Indicative price (Apple Store, August 27, 2026): M5 Max from $3,300 (36 GB) up to 128 GB; M5 Ultra from $7,200 (96 GB), $13,600 in 256 GB, 512 GB price not yet announced.
• Memory bandwidth: 614 GB/s (M5 Max), 1.2 TB/s (M5 Ultra) — this, not the GPU core count, is what caps generation speed.
• Upgradeability: proprietary. RAM is soldered onto the SoC, SSD is soldered, no user-replaceable components. You pick the configuration at purchase, for the lifetime of the machine.
• Software: macOS + MLX, LM Studio, Ollama. No native GNU/Linux.

Mind the trap of the base M5 Max: 36 GB of memory is quickly a wall for 70B-class models even though the GPU and bandwidth are excellent. The 128 GB configuration is the one that makes sense for serious local AI. Note also that with quantized frontier-class MoE models (around 170 GB in 4-bit for a 320B model), a single Studio is no longer enough — Apple pitches Thunderbolt 5 clustering of several machines to load them.

Apple Mac mini (M6 / M5 Pro)

The Mac mini is the Studio's little sibling: same Apple unified-memory logic, but in an even smaller case and with a more modest spec sheet. The latest M6 model, announced on August 25, 2026, starts at $1,150 with 16 GB of unified memory and can be configured up to 32 GB ($220 more for 24–32 GB, which also unlocks the full 170 GB/s bandwidth — the 16 GB version is capped at 153 GB/s). It comfortably runs 7B-14B models and even a quantized 32B. Above it, the Mac mini M5 Pro starts at $2,200 and goes up to 64 GB of unified memory, opening the door to quantized 70B models. It's Apple's entry-level option for anyone who wants to taste local AI without investing in a Studio.

• Maximum memory: 32 GB (M6) / 64 GB (M5 Pro), unified memory.
• Indicative price (Apple Store, August 27, 2026): M6 from $1,150 (16 GB), 24–32 GB configurations from around $1,370; M5 Pro from $2,200 (up to 64 GB).
• Memory bandwidth: 153 GB/s (M6 16 GB), 170 GB/s (M6 24–32 GB), 307 GB/s (M5 Pro).
• Upgradeability: proprietary. Soldered RAM, soldered SSD, no user-replaceable components.
• Software: macOS + MLX, LM Studio, Ollama. No native GNU/Linux.

This is the simplest choice to start with local AI on Apple hardware — prefer the 24 or 32 GB configuration over the 16 GB base model, the extra memory is what decides which models you can load. Note the M6 starts $380 more than the M4 it replaces, a consequence of the ongoing RAM price crisis, and Apple has announced no M6 Pro or M6 Max (the next step would be an M7 expected around mid-2027).

Nvidia DGX Spark (Dell)

Nvidia launched with Dell the DGX Spark, a compact workstation built for developers who want a local AI node with the full CUDA stack. The Spark reaches 128 GB of unified memory on the Grace Hopper / Superchip platform and ships with Nvidia's software suite (NIM, cuOpt, TensorRT), positioning it as a mini prototyping system for GPU workloads.

• Maximum memory: 128 GB (unified memory).
• Indicative price: around $4,300 (standard configuration).
• Upgradeability: proprietary. The machine ships assembled and closed; RAM is integrated into the SoC.
• Software: Nvidia DGX OS (based on Ubuntu/GNU/Linux), CUDA, NIM, Ollama. Compatible with the full AI development ecosystem.

This is the option to consider if your workflow relies on CUDA, TensorRT or NIM microservices and you want an officially supported Nvidia environment, in a desktop form factor.

Microsoft Surface mini (Copilot+)

Microsoft positioned itself on the niche with a mini Surface from the Copilot+ family: compact form factor, Snapdragon X chip (or x86 equivalent depending on the variant), able — per Microsoft — to run models up to 120 billion parameters via the NPU emphasis and aggressive quantization. In practice, this kind of machine handles 7B-14B models locally and smoothly, with larger models working in "chunked loading" mode.

• Maximum memory: 64 GB (unified memory, depending on configuration).
• Indicative price: $1,600–$2,700 depending on RAM and storage.
• Upgradeability: proprietary. Soldered RAM, soldered SSD on most Surface variants.
• Software: Windows 11 Copilot+ (NPU enabled), with possible GNU/Linux dual-boot with caveats (Snapdragon support still uneven).

This is a "consumer local AI" option: silent, compact, integrated with Windows, but closed on the hardware side.

Minisforum (AI lineup / UM780 XTX)

Minisforum goes head-on against the Mac mini with a lineup of mini-PCs that accept standard DDR5 SO-DIMM, making them the most upgradable machines in this comparison. The UM780 XTX reaches 64 GB (or up to 96 GB on the newer Strix Halo SoCs), and the recent AI lineup pushes AMD unified memory further.

• Maximum memory: 64–96 GB depending on the SoC (replaceable DDR5 SO-DIMM).
• Indicative price: $750–$1,600 depending on the SoC and configuration.
• Upgradeability: semi-upgradable. RAM is replaceable by opening the case (SO-DIMM), the NVMe SSD too. The GPU is integrated into the SoC and not replaceable.
• Software: GNU/Linux (Debian, Fedora, NixOS) supported without issue, Windows 11 too. Ollama, ComfyUI, llama.cpp work via ROCm 6.5+ on AMD SoCs.

This is the most libre-friendly choice of the comparison: Mac mini form factor, but open hardware and GNU/Linux-friendly, for anyone who wants to upgrade RAM themselves.

Xiaomi AI Cube

Xiaomi unveiled its AI Cube, a mini-PC with an in-house chip (Xiaomi / MediaTek SoC per rumors), positioned on the Asian market first. Few official details on the full spec sheet, but the stated goal is a compact, affordable device built to run models locally, leaning on the in-house NPU.

• Maximum memory: 32 GB (to be confirmed, spec sheet not finalized).
• Indicative price: $550–$1,000 on the Chinese market (European availability to be confirmed).
• Upgradeability: proprietary. In-house chip, RAM likely soldered, no official opening.
• Software: Xiaomi skin on top of embedded Android/Linux; compatibility with the open-source AI stack (Ollama, llama.cpp) to be verified case by case.

This is the machine to watch for the price/performance ratio on the Asian market, but the software closedness makes it a gamble for anyone wanting a fully open-source AI stack.

Framework Desktop 128 GB

Framework — known for its modular laptops — launched the Framework Desktop, a 4.5L Mini-ITX desktop powered by AMD Ryzen AI Max+ 395 (Strix Halo), supporting up to 128 GB of unified LPDDR5x memory. To our knowledge, this is the first machine in this niche to reach 128 GB of unified memory without going through Apple. The machine is already available (first shipments in Q3 2025). As Framework officially explains, the RAM is soldered onto the SoC: LPDDR5x is imposed by Strix Halo's 256-bit memory bus to reach 256 GB/s of bandwidth, and Framework determined that modular RAM at that throughput was not technically feasible. On the other hand, the NVMe SSD, the power supply (Flex ATX), the fan (standard 120 mm) and the Framework expansion modules remain replaceable.

• Maximum memory: 128 GB (unified LPDDR5x memory, soldered).
• Indicative price: Ryzen AI Max+ 395 / 128 GB configuration starting at $1,999; base configuration (Ryzen AI Max 385 / 32 GB) starting at $1,099.
• Upgradeability: semi-upgradable. Soldered RAM (not replaceable), replaceable NVMe SSD (2 M.2 2280 slots, up to 16 TB), standard fan and power supply, Framework modules.
• Software: GNU/Linux natively supported (Framework pushes Fedora and Ubuntu, Bazzite/Playtron possible), Windows 11, Ollama via ROCm on the AMD SoC.

This is the most credible alternative to the Mac Studio for anyone who wants 128 GB of unified memory with an open SSD and GNU/Linux-friendly, accepting that the RAM is soldered.

Upgradable gaming PC (RTX 5090)

A custom-built gaming PC remains the most hardware-libre option in this comparison: everything is replaceable. With an RTX 5090 (32 GB VRAM) and 128 GB of DDR5, you have enough to run quantized 30B-70B models comfortably, with the mature CUDA stack for Ollama, ComfyUI and llama.cpp. It's also the only option where you can increase VRAM later by adding a second card (dual-GPU), doubling model capacity without changing machines.

• Maximum memory: 128 GB DDR5 system + 32 GB VRAM (up to 96 GB in dual-GPU).
• Indicative price: $2,700–$4,800 depending on GPU, CPU and RAM.
• Upgradeability: upgradable. RAM, GPU, SSD, PSU and motherboard are replaceable at will. This is the only machine in the comparison where VRAM can scale up via a second GPU.
• Software: GNU/Linux (Debian, Fedora, NixOS) natively supported, Windows 11 too. Ollama, ComfyUI, llama.cpp via CUDA; ROCm possible if you switch to AMD.

This is the choice for anyone who wants to upgrade their AI hardware over time, accepts a larger case and higher power draw (budget for a 1000W PSU for an RTX 5090).