What's the best mini PC for local AI, and how much memory do you need?

Short answer: the best mini PC for local AI is the cheapest one with enough unified memory for the biggest model you'll run. Memory decides what loads; the brand mostly decides speed and software. On our VRAM engine at 4-bit, 32GB runs 27B models, 64GB runs 30B to 35B models at near-lossless 8-bit (and a 70B only tight), 128GB runs gpt-oss-120b and Qwen3.5 122B, and 235B models need 256GB. For 128GB, the cheapest box we found on November 6, 2026 is an AMD Strix Halo mini PC; for 64GB and under, it's a Mac mini.

How much unified memory do you need for local AI?

Enough to hold the model, its context and the operating system, with room left over. On a mini PC the CPU and GPU share one pool of memory, but the GPU only gets part of it. Macs give the GPU roughly two-thirds to three-quarters (see below), and on Strix Halo you set the GPU's share in the BIOS. So size the memory for your biggest model, then add a margin.

Here's what our VRAM calculator engine says popular open models need at 4-bit (Q4_K_M) with a 16K context, computed November 6, 2026:

The GPU share column is our planning figure: two-thirds of memory at 32GB and below, three-quarters above. That's what we use for Macs, and on a 128GB Strix Halo it matches AMD's own 96GB setting. Check your exact model, quantization and context length in the VRAM calculator; a longer context needs more memory than these totals. If you're unsure between 4-bit and 8-bit, our quantization guide explains the trade.

Two rules fall out of the table. First, you can't add memory later: none of these machines has memory slots, and Framework lists its Desktop memory as "non-upgradeable". Buy for the model you'll want next year. Second, 64GB is the awkward tier for 70B models. A 70B at 4-bit uses about 95% of a 64GB machine's GPU share on our math. It may load, but leaves little room for a long context. If 70B is the goal, buy 128GB.

What's the best mini PC for local AI right now?

It depends on the memory tier you landed on. These are the boxes we'd pick, with each maker's US price as we read it.

Apple's configurator prices each option as a difference from the model on screen. On November 6, the 32GB Mac mini M6 with 1TB was $1,799 and 256GB storage was $500 less; the 18-core M5 Pro with 64GB and 1TB was $3,199, the 15-core chip was $200 less and 512GB storage $300 less; the 40-core M5 Max Mac Studio with 64GB and 1TB was $3,799, with 128GB adding $1,600; and the M5 Ultra with 96GB and 1TB was $5,499, with 256GB adding $4,000.

Strix Halo: the cheapest way to 128GB

AMD's Ryzen AI Max+ 395, nicknamed Strix Halo, is a laptop-class chip with a large built-in GPU and up to 128GB of LPDDR5x-8000 on a 256-bit bus, per AMD's spec page (read Nov 3, 2026). That works out to about 256 GB/s. Several brands sell it in mini PCs; we priced GMKtec's EVO-X2 and the Framework Desktop because both list prices on their own stores. On November 6, GMKtec listed the 128GB, 1TB EVO-X2 at $3,499.99 on sale ($3,999.99 regular) and the 2TB version at $3,649.99.

If you saw "about $1,999" for a 128GB Strix Halo box earlier this year, that price is gone. An earlier version of this page said it too. Memory prices have pushed these machines up across 2026, so check the date on any price you read. Framework's 128GB Desktop DIY Edition was $3,449 when we read its US store on November 3. On November 6 we could only load Framework's European store, where every Ryzen AI Max option was marked "Out of stock".

On a 128GB Strix Halo, AMD's Variable Graphics Memory gives the GPU up to 96GB, and AMD's FAQ (read Nov 3, 2026) says the GPU "can technically access a total graphics memory size of 112GB". Our Strix Halo VRAM guide explains the settings, and a hands-on run of Qwen models on one shows what it's like in use.

A newer chip is arriving: Framework opened pre-orders on September 30, 2026 for a 192GB Framework Desktop with the Ryzen AI Max+ PRO 495, "with shipments starting in November". Framework says it is "designed specifically for Linux" and that "neither the DIY Edition nor the pre-built versions have Windows options available". Framework's store listed it at $6,799 (DIY) when we read it on November 3.

Mac mini and Mac Studio: faster replies, more money per gigabyte

Macs with Pro, Max and Ultra chips generate text faster than Strix Halo at the same memory size because their memory is faster; the plain M6 is slower. Apple's spec pages, read November 6, list 153 GB/s for the two 16GB Mac mini M6 models and 170 GB/s for the 24GB one, 307 GB/s for the M5 Pro, 460 GB/s for the 32-core M5 Max, 614 GB/s for the 40-core M5 Max and 1.2 TB/s for the M5 Ultra.

The catch is the ceiling. The Mac mini stops at 64GB (the M6 at 32GB), so a 128GB Mac means a Mac Studio, and that costs about $1,600 more than the cheapest 128GB Strix Halo box. The M5 Ultra Mac Studio is the only mini PC we found that goes past 192GB: 256GB with the standard chip, or 512GB with the 80-core chip. For a full config-by-config breakdown, see our Mac mini buying guide and the best Mac for local AI.

DGX Spark: buy it for CUDA, not for value

NVIDIA's DGX Spark uses the GB10 chip with 64GB or 128GB of LPDDR5x, which NVIDIA calls "coherent unified system memory", and 273 GB/s, per NVIDIA's product page on November 6. NVIDIA lists one 128GB unit for models "Up to 200B parameters" and says the 64GB version "supports up to 100-billion-parameter models".

Its price has moved all year. NVIDIA's product page shows no price. On February 25, 2026, NVIDIA announced on its developer forum that the Founders Edition MSRP had moved from $3,999 to $4,699, citing "worldwide constraints in memory supply". In October, AI Weekly, citing Tom's Hardware, reported that "Nvidia's 128GB DGX Spark now lists at $6,950". The 64GB version went on sale from Acer, ASUS, Dell, Gigabyte, HP and MSI on October 23, "starting at $4,999", per NVIDIA's blog.

Bandwidth-wise it sits with Strix Halo, so replies come at a similar pace. What you pay for is NVIDIA's software stack, faster prompt processing and better image-generation support. Our three-way comparison has the benchmarks, and our DGX Spark review covers who it's for.

Why doesn't the GPU get all the memory?

Because the operating system and your other apps need some, and each platform caps the GPU's share. On a Mac, macOS sets a recommended GPU working-set limit (Apple's Metal documentation calls it recommendedMaxWorkingSetSize). Apple doesn't publish the percentage, so we plan on two-thirds of memory at 32GB and below and three-quarters above, based on values llama.cpp users have logged. Our Mac mini guide shows those logs. On a 128GB Strix Halo, AMD's own settings top out at 96GB for the GPU. On the DGX Spark, the memory is "coherent unified system memory" shared by the CPU and GPU, and the OS still needs its part.

That's why "a 128GB mini PC" doesn't mean "a 128GB model". Plan on about 96GB for the model and its context.

Is a mini PC faster than a graphics card for local AI?

Not on models that fit both. Writing a reply means reading the model's active weights from memory for every token, so speed follows memory bandwidth, and a desktop graphics card has several times the bandwidth of a 256 to 273 GB/s mini PC. A mini PC wins on a different question: it can hold a 70B to 120B model that no single consumer card can.

Two things soften the speed gap. Many of the best large open models are mixture-of-experts models (gpt-oss-120b, Qwen3.5 122B-A10B, Llama 4 Scout), which read only a fraction of their weights per token, so they stay usable on slower memory. And these boxes are small, quiet and draw far less power than a multi-GPU tower.

If every model you run fits in 24 to 32GB, skip the mini PC and buy a graphics card; our best GPU for local AI guide covers which. And if you're not sure yet, rent one by the hour first.

Should you rent before you buy a mini PC?

For the DGX Spark, yes: you can rent its chip. On November 6, 2026 at 21:20 UTC our price tracker read one GB10 machine on Vast at $0.28 an hour on demand, so an evening of testing your own models and prompt lengths costs a few dollars against a purchase in the thousands. Prices move hourly; our cloud GPU page has today's, and cost to run gpt-oss-120b shows the cheapest rental that fits that model.

Macs and Strix Halo boxes aren't rentable the same way, so for those, published benchmarks are what you have. Our rent vs buy guide works through when owning pays off.

Unified memory is one half of the local-AI hardware story; graphics cards are the other. For the GPU side, see the best GPU for local AI, and for one model that shows why these boxes matter, see what gpt-oss-120b needs to run.