Personal AI computers compared: what each one actually runs

A personal AI computer is a machine that runs open models at home, all day, without sending anything to a cloud. Two numbers decide what one can do: how much memory its GPU can use, which sets the biggest model it holds, and how fast the GPU reads that memory, which sets how quickly answers come. Everything below is computed from the makers' published specs with the same formulas as the rest of this site. It is not a hands-on review.

The numbers that matter

"GPU can use" is the figure every fit on this site is measured against. On the DGX Spark it is what the system leaves free, less a 2 GB display reserve; on an RTX Spark it is the 102 GB NVIDIA's memory rule guarantees on a 128 GB machine, since NVIDIA does not print the size of the block it reserves for the GPU; on the Ryzen AI Max+ 395 it is the 96 GB AMD lets you set aside as graphics memory; on the Mac it is about 75% of RAM, the macOS limit observed on large-memory Macs. Each machine is shown at its largest memory option; the 512 GB Mac Studio M5 Ultra ships in late October 2026. The machines that share memory hold far more, and read it more slowly, than the ones with a graphics card. The DGX Spark, Strix Halo and Mac Studio comparison goes deeper on those three.

What each one runs

Of the 34 open models tracked on this site, at Q4_K_M with 8k of context, and for an agent at the 65,536 tokens of context Ollama recommends for OpenClaw:

"The one to run" is the same pick as on each GPU page: the biggest class that fits with room for a conversation and generates at a usable speed, then the newest release. The agent column uses the rule from the OpenClaw guide: only models whose chat template passes tools through, with the context already counted.

Speed on the same models

Estimated tokens per second at Q4_K_M and 8k of context, single stream. These are ceilings worked out from memory bandwidth, not measurements; figures marked ≤ are where real runtimes fall furthest below them. "No" means the model does not fit.

On a model both can hold, the card with the faster memory wins: the RTX 5090 and the RTX PRO 6000 read memory at 1,792 GB/s, and the Ghost Core's card at 432 GB/s is still 1.6 times as fast as the DGX Spark. The machines with a large shared pool win as soon as the model stops fitting on a card. These figures describe generation; reading a long prompt first is a separate cost, and it is markedly slower on Apple silicon than on NVIDIA hardware.

Ghost Core, up close

Ghost's spec page lists an NVIDIA RTX PRO 4000 Blackwell with 24 GB, a 6-core Ryzen 5 7600, 64 GB of DDR5 and a 1 TB SSD, and says the first batch ships on 31 October 2026. It also lists 432 GB/s of memory bandwidth and 770 AI TOPS. Those are the figures NVIDIA prints for the SFF Edition of the card, a 70 W version, not for the full-size RTX PRO 4000, which reads its memory at 672 GB/s. This page uses the SFF figures.

Ghost lists four models. Three are on this site's list:

The fourth, "Qwen 3.8-Next", matches no model on Hugging Face by that name. The nearest is Qwen3.8-Flash-Next: 125B parameters with 6B active, plus a 51B n-gram embedding that Qwen says suits offloading. On a 24 GB card most of that would have to live in the 64 GB of system memory. This site computes GPU memory only, so it gives no figures for that model; ask Ghost which version and quantisation it runs.

On this card those three generate at about 17 to 19 tokens per second at best, below the 25 this site calls comfortable for chat. That is why the everyday pick in the table above is the smaller Gemma 4 12B.

For an agent, the card holds what it needs. Muse Glimmer 30B, Meta's model built for local agents, needs 19.4 GB at 64k of context, against the 22.8 GB this site plans for on a 24 GB card. Qwen3.8 27B needs 21.7 GB at the same setting, at about 19 tokens per second. The card cannot hold a 70B model: Llama 3.3 70B needs about 45.8 GB at Q4 with 8k of context.

Which to choose

• Ghost Core if you want a quiet box that runs a 27B to 30B model all day and gives your agents a local endpoint. It holds the least memory here, and its card is the slowest of the three graphics cards on this page.
• DGX Spark if you need CUDA and room for 70B and 120B-class models. It reads memory at 273 GB/s, so big models generate slowly.
• An RTX Spark laptop if you want CUDA and room for 70B and 120B-class models in a laptop, on sale from 16 October 2026. It reads memory at 300 GB/s, so big models generate slowly, and only the 128 GB configurations hold them.
• A Ryzen AI Max+ 395 mini PC if you want one x86 machine that runs Windows and holds large mixture-of-experts models. Check the power limit of the specific machine.
• A Mac Studio M5 Ultra if you want the most memory and the fastest generation of the shared-memory machines, and do not need CUDA.
• A PC with an RTX 5090 if speed matters most and your models fit in 32 GB.
• An RTX PRO 6000 if you want 96 GB on one card at the 5090's speed, in a workstation.

Or no hardware at all

None of these is the cheap way to run a big model now and then. If the large models, a long video render or a fine-tune are an occasional need, renting a larger GPU for those hours costs less than a machine sized for them. The build vs rent calculator finds the break-even month with your own prices and hours.

To rent one, the cloud GPU price guide lists the cheapest hour for each card, and ComfyUI in the cloud compares the services that bill per generation instead.

Questions

What is the best personal AI computer?

It depends on the biggest model you want to run. For the largest models, a Mac Studio M5 Ultra: its GPU can use up to 384 GB. For CUDA with a lot of memory, the DGX Spark (126 GB), or an RTX Spark laptop for the same idea on the move (102 GB). For speed on models up to 32 GB, a PC with an RTX 5090, which reads memory at 1,792 GB/s. For a small, quiet box that runs 27B to 30B models all day, the Ghost Core.

Can the Ghost Core run OpenClaw?

Yes. At the 64k of context Ollama recommends for OpenClaw, Qwen3.8 27B fits, but tightly: it needs 21.7 GB of the 22.8 GB this site plans for on the card, at about 19 tokens per second. Muse Glimmer 30B, which Ghost also lists, fits with room at 19.4 GB.

Ghost Core or DGX Spark?

The Ghost Core reads memory 1.6 times as fast (432 GB/s against 273 GB/s), so it generates faster on any model both can hold. The DGX Spark gives its GPU 126 GB against 24, about 5 times as much, which is what it takes to run 70B and 120B-class models. Both use NVIDIA GPUs.

Can a personal AI computer run a 70B model?

At Q4_K_M with 8k of context, Llama 3.3 70B needs about 45.8 GB. That fits on the DGX Spark, an RTX Spark laptop, a Ryzen AI Max+ 395 mini PC, a Mac Studio M5 Ultra and an RTX PRO 6000 workstation. It does not fit on a 24 or 32 GB card.

Do I need a personal AI computer, or can I rent one?

Buying makes sense when the machine is busy most days. If the big models are an occasional need, renting a larger GPU for those hours costs less than a machine sized for them; the build vs rent calculator finds the break-even with your own numbers.

• Ghost Core: Ghost's product page, read on 6 October 2026. Card figures from NVIDIA's RTX PRO 4000 Blackwell SFF and RTX PRO 4000 Blackwell pages.
• RTX Spark: NVIDIA's product page for the memory, and its developer porting guide and unified-memory guide for the bandwidth and the GPU's share, read on 9 October 2026.
• Memory and bandwidth for every other machine: each maker's specification, linked on each machine's GPU page.
• Models: each model's config.json; Qwen3.8-Flash-Next's parameter counts from its model card.
• Fits and speeds: the same formulas as the VRAM calculator and the speed estimator.