DGX Spark, Ryzen AI Max or Mac Studio: which small AI workstation makes sense?

A new kind of computer has appeared for people who want to run large AI models at home or in a small office. It replaces the usual separation between main memory and graphics memory with one shared pool. That one change lets a small box load a 65 GB model, which a 32 GB gaming card selling for more than US$8,000 cannot hold.

The catch is speed. Shared memory is much slower than the memory on a dedicated card, and LMSYS's review names it as the main bottleneck. These machines are about capacity first, and the differences between them come down to bandwidth, software and price.

The three families

NVIDIA DGX Spark uses the GB10 chip, a 20-core Arm processor joined to a Blackwell graphics processor with 6,144 CUDA cores. It runs DGX OS, a Linux system with NVIDIA's AI software preinstalled, and its two ConnectX-7 ports let you link units together.

AMD Ryzen AI Max, known by the code name Strix Halo, is sold by many makers, including Framework and GMKtec. It is an ordinary x86 PC that runs Windows or Linux and can also play games. On Windows, you set aside up to 96 GB of a 128 GB system as graphics memory, after a restart.

Apple Mac Studio uses M5 Max or M5 Ultra chips and is the only one of the three that goes up to 512 GB.

A fourth option arrived in October: RTX Spark laptops and compact desktops running Windows, built around the N1X, a close relative of the DGX Spark's chip. StorageReview reports up to 128 GB, with laptops running at 45 to 80 W and desktops at 140 W. Independent speed tests are still thin.

Specifications and prices

Prices were read on November 2, 2026, except where noted, and exclude tax.

The DGX Spark prices are NVIDIA's, as reported in October by Implicator and in NVIDIA's own blog for the 64 GB model. The 128 GB unit had launched at US$3,999 a year earlier.

Clustering changes the arithmetic. Two 64 GB DGX Spark units pool to 128 GB and double the memory bandwidth, and NVIDIA reports up to 1.7 times the performance of one unit on its own Qwen 3.8 27B test. That costs at least US$9,998 for the pair, against US$6,950 for one 128 GB machine, and no independent test had confirmed the figure when Implicator reported it.

What the tests show

Tom's Hardware put a DGX Spark against a Ryzen AI Max+ 395 machine in January 2026. The Spark was much faster at reading long prompts, so the first word arrives sooner, especially at long context. It also led on writing speed, except on dense Gemma 3 27B, where the two were close. In one image generation workflow, the AMD machine took about four times as long. Tom's also hit errors on a video workflow on AMD's ROCm software, a reminder that NVIDIA's software is more mature.

For a sense of absolute speed, LMSYS measured the Spark at 49.7 tokens per second on gpt-oss-20b, against 205 for an RTX 5090. On the Apple side, the llama.cpp project's public benchmark table shows an M5 Ultra about 50% faster than an M5 Max on a small test model. No single test puts all of these on equal terms, so treat comparisons across different sources as rough.

A worked example: picking one for a small team

Imagine an office of five that wants a shared assistant for private documents, running Qwen3.8-27B (18 GB).

• Ryzen AI Max+ with 64 GB: the model fits with room for long context, and it is the cheapest route. It is a good choice if a few short waits are acceptable.
• DGX Spark with 64 GB: it costs more but reads long documents much faster, and LMSYS found its batching strong, with total throughput on Llama 3.1 8B rising from 20.5 to 368 tokens per second between one and 32 simultaneous requests.
• Mac Studio with 64 GB: it has the highest bandwidth of the three and low power use, though vLLM support on Mac depends on a community plugin. See the server guide.

For this office, the DGX Spark's batching makes a strong case despite its price. One person who mostly chats might reasonably decide otherwise.

Is a small workstation enough to replace the cloud?

For many everyday tasks, yes, within limits. Framework's September 2026 tests show a 128 GB machine loading DeepSeek V4 Flash in 90.9 GB at about 2 bits per weight, and that compression is heavy. So you can run a very large model at home, but in a much reduced form, and slowly compared with a data centre. The largest open models, and any closed frontier model, stay out of reach.

Some owners have also reported shutdowns under sustained load on DGX Spark units in NVIDIA's developer forums. In that thread a staff member calls it a known issue and suggests lowering the maximum GPU clock. These are individual cases, not a measured failure rate, but they are a reason to buy from a seller with a clear return policy.

So treat a workstation as one part of a mix: private and routine work stays local, and hard questions go to a hosted model, as the cloud or local guide lays out.