View all articles
HardwareLocal AIBuying GuideVRAM

Local AI hardware 2026: what to buy, priced and sized

JG
Jacobo González Jaspe
|

Reviewed:

Abstract illustration: thin glass layers stacked like a chip, with an amber core on a navy background. Hardware
Illustration generated with AI on our own machine.
This article is also available in Spanish:Hardware para IA local 2026: qué comprar, con precio y medida

Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you. It does not change what we recommend. How this works

By the end of this guide you will know which machine holds the model you need, what it cost in Spain on 16 November 2026, and whether you should buy anything at all. A calculator and the number of people who will use it are enough to follow along.

Who this is for, and when to buy nothing

This is for a small company or a tinkerer choosing a machine for local language models, run with Ollama or llama.cpp.

Buying nothing is often right. If your text holds no personal or confidential data, or you need the largest models, a cloud API in a European region is the simpler answer: the levels and their contracts are in local AI: what it is and when it pays off. And if you have never run a model, try the laptop you own first: an 8B needs about 6.5 GB.

How these prices were collected

Every price below was read on 16 November 2026 from the page linked next to it: the maker’s own Spanish store where it has one, otherwise a named retailer’s product page. All are new units unless the row says second-hand. Card prices are for the cheapest model we found sold by the retailer itself, not by a marketplace seller.

This is a snapshot of one day. Prices moved between our September version and this one: the 16 GB card we listed at EUR 550 to 700 now starts at EUR 859.99. To recheck a row, open its link. If the link is dead, treat the row as gone. Anything we could not source that day is not here, so every September price has been replaced or removed.

The catalogue: graphics cards for a PC you own

You need a desktop with a free PCIe slot and the power supply the maker asks for.

CardMemoryMaker’s specPrice, 16 Nov 2026
Intel Arc B58012 GB190 W, 600 W supply, 456 GB/s (Intel)EUR 470.13, ASRock Challenger at PcComponentes
NVIDIA RTX 306012 GB170 W, 550 W supply (NVIDIA)EUR 509.00, Gigabyte Windforce OC at PcComponentes
NVIDIA RTX 5060 Ti16 GB180 W, 600 W supply (NVIDIA)EUR 859.99, PNY Overclocked Dual Fan at PcComponentes
NVIDIA RTX 3090, second-hand24 GB350 W, 750 W supply (NVIDIA)EUR 900 to 1,400 asked, private sellers on Wallapop
Intel Arc Pro B7032 GB230 W, 608 GB/s (Intel)EUR 2,058.49, ASRock Creator at PcComponentes
NVIDIA RTX 509032 GB575 W, 1,000 W supply (NVIDIA)EUR 9,999.99, MSI Gaming Trio OC at PcComponentes

Notes on two rows:

  • RTX 3090. The range is what the eleven cheapest listings we read asked, with a EUR 400 floor to drop accessories. An asking price is not a sale price and not a list price. No shop we could open sold one used: CeX Spain had none in stock, and its nearest card was a used RTX 3080 Ti 12 GB at EUR 590. Sold as new by third parties, the RTX 3090 on Amazon.es was EUR 2,989.99.
  • RTX 5090. Marketplace sellers on the same page started at EUR 8,521.50. The NVIDIA pages we opened showed no price for any card.

The catalogue: complete machines

Unified memory is one pool shared by processor and graphics chip. Bandwidth is how fast it is read, and it caps speed (see below).

MachineMemory and storageMaker’s specPrice, 16 Nov 2026
Mac mini M616 GB, 256 GB153 GB/s, 155 W maximum (Apple)EUR 1,069, Apple; EUR 979.90 at PcComponentes
Mac mini M624 GB, 512 GB170 GB/sEUR 1,509, Apple
Mac mini M632 GB, 1 TB170 GB/sEUR 2,059, Apple
Mac mini M5 Pro, 18-core64 GB, 1 TB307 GB/sEUR 3,669, Apple
Mac Studio M5 Max36 GB, 512 GB460 GB/s, 480 W maximum (Apple)EUR 3,029, Apple
Mac Studio M5 Max, 40-core GPU64 GB, 1 TB614 GB/sEUR 4,459, Apple
Mac Studio M5 Ultra96 GB, 1 TB1.2 TB/sEUR 6,649, Apple
Framework Desktop, Ryzen AI Max+ 39564 or 128 GB, no drivememory not upgradeable; no bandwidth on the pageEUR 2,239 or 3,919, Framework, out of stock
Gigabyte AI TOP ATOM (NVIDIA GB10)128 GB, 4 TBsame chip as the DGX SparkEUR 6,999.90, PcComponentes
NVIDIA DGX Spark Founders Edition128 GB, 4 TB273 GB/s, 240 W supply (NVIDIA)EUR 7,490, NVIDIA Marketplace
NVIDIA Jetson Orin Nano Super8 GB, SD card102 GB/s, 7 to 25 W (NVIDIA)USD 249 on NVIDIA’s page; no euro price from NVIDIA

The Apple prices are standard configurations in the store’s configurator, VAT included. The Framework price has no drive, operating system or power cable: each is an extra on the same page.

The Jetson is the one row where the maker’s price and the street disagree. On Amazon.es, a listing titled “NVIDIA Jetson Orin Nano Development Kit” was EUR 458.69, shipped by Amazon US, with one unit left: see the Jetson Orin Nano on Amazon.es.

What each machine can hold

The method is the one in our VRAM guide, where the sum matched 13 real loads to the megabyte (load-log receipt):

text
memory = weights (the file) + context cache × parallel slots + compute buffers
cache per token (bytes) = layers × KV heads × head dimension × 2 × 2

We worked it for five models at Q4_K_M, the usual quantisation. Buffers are an allowance: 0.5 GB up to 14B, 1 GB above. The sum loads nothing. Columns, in GB: 8k of context for one person, 32k for one person, 8k for four people at once.

text
8B  Llama 3.1 8B        6.49    9.71    9.71
14B Qwen2.5 14B        11.10   15.93   15.93
24B Mistral Small      16.67   20.70   20.70
32B Qwen2.5 32B        23.00   29.44   29.44
70B Llama 3.3 70B      46.20   54.26   54.26

The last two columns are equal: four people at 8k reserve the same cache as one person at 32k. For the 14B and 32B, file sizes are from bartowski’s repositories (14B, 32B) and layer counts from each config.json (14B, 32B), read on 2026-11-16. The rest come from the VRAM guide and the Llama 3.3 70B guide.

A check against our own loads: at 32k the sum gives 9.71 GB for the 8B and 15.93 GB for a 14B. On our first Spark, Ollama reported 9.21 GB for llama3.1:8b and 15.29 GB for deepseek-r1:14b, which has the same file size and layers (receipt). The sum errs about half a gigabyte high, the safe side.

Now the machines. As in the VRAM guide, a card keeps 0.5 GB back. Unified memory keeps 3 GB for the system up to 16 GB, 4 GB up to 32 GB and 6 GB above. “Tight” means over 85% of what is left. Each cell reads “8k for one person / 32k for one person or 8k for four”; a single “no” means both.

MemoryUsable8B14B24B32B70B
8 GB unified (Jetson)5 GBnonononono
12 GB card (B580, 3060)11.5 GByes/yestight/nononono
16 GB card (5060 Ti)15.5 GByes/yesyes/nononono
16 GB unified (Mac mini M6)13 GByes/yestight/nononono
24 GB card (3090)23.5 GByes/yesyes/yesyes/tighttight/nono
24 GB unified (Mac mini)20 GByes/yesyes/yesyes/nonono
32 GB card (B70, 5090)31.5 GByes/yesyes/yesyes/yesyes/tightno
32 GB unified (Mac mini M6)28 GByes/yesyes/yesyes/yesyes/nono
36 GB unified (Mac Studio)30 GByes/yesyes/yesyes/yesyes/tightno
Two 24 GB cards47 GByes/yesyes/yesyes/yesyes/yestight/no
64 GB unified58 GByes/yesyes/yesyes/yesyes/yesyes/tight
96 or 128 GB unified90, 122 GByes/yesyes/yesyes/yesyes/yesyes/yes

Four things the table cannot show:

  • A card’s memory is its own; unified memory is shared. On a card, the whole sum must fit in VRAM, and the PC’s own RAM is not counted. On unified memory the model competes with the system and every other program. Our “128 GB” machines report 121.7 GB in total, and had 65.4 to 75.1 GB available before the loads in our receipts.
  • macOS decides how much the graphics chip may hold. Apple publishes no fixed share of RAM for it, so the Mac rows are an upper bound. The reserve we subtract is our planning convention, not Apple’s figure.
  • Two cards are not one pool. Whole layers go to one card or the other, so expect less room than 47 GB. We have not run a split load.
  • The 32B on 24 GB is the edge case. The VRAM guide’s table uses a more cautious sum and puts it out by 0.03 GB; this one puts it in with 0.5 GB spare. Count it as the limit.

The 8 GB Jetson holds none of these five: it is a machine for smaller models, which we did not work here. For a first shortlist by memory, the site has a VRAM explorer. Which file to pick when the usual one does not fit is in how to pick the right GGUF file.

How fast: what we measured, what others published

Measured by us, on our two DGX Sparks, with the protocol of our 30-minute benchmark guide: one 278-word prompt, 200 tokens out, second run reported.

Model (Q4_K_M)FileGenerationReceipt
llama3.1:8b, dense4.92 GB43.7 tok/s2026-11-02
deepseek-r1:14b, dense, thinking off8.99 GB22.4 tok/ssame file
mistral-small:24b, dense14.33 GB14.4 tok/s2026-11-15
gemma4:26b, mixture of experts17.99 GB68.2 tok/s2026-11-02
qwen3.6:35b, mixture of experts23.94 GB75.8 tok/ssame file

More in our DGX Spark review and in Mistral Small 24B, measured.

Published by others. Two sources we opened on 2026-11-16:

HardwareModel at Q4_K_MGenerationSource and conditions
DGX Sparkllama3.1 8B38.02 tok/sOllama’s blog, 23 Oct 2025, Ollama 0.12.6, 500 tokens out
DGX Sparkqwen3 32B9.41 tok/ssame
DGX Sparkllama3.1 70B4.42 tok/ssame
RTX 3090, 24 GBLlama 3 8B111.74 tok/sXiongjieDai’s benchmark repository, May 2024, llama.cpp, 1,024 tokens, rented machines
Two RTX 3090Llama 3 70B16.29 tok/ssame

For every other machine in the catalogue we make no speed claim. We found no measurement of the M6 or M5 Macs, the RTX 50 cards, the Arc cards or the Framework Desktop that names its software and settings, and we own none of them. The third-party speeds in the September version are removed.

What you can work out without a benchmark is a ceiling. A dense model reads all its weights for each token, so bandwidth divided by file size caps the speed for one person. On our Spark that is 273 ÷ 14.33 = 19.1 tokens per second for the 24B, and we measured 14.4; for the 8B, 55.5 against 43.7. Our machine reaches about three quarters of its ceiling. Whether a Mac does the same we cannot say. Its ceiling for that 24B file is 11.9 tokens per second at 170 GB/s, 21.4 at 307 and 32.1 at 460: sums, not measurements.

Mixture-of-experts models break the ceiling because they read only part of their weights per token. That is why our two fastest rows are our two largest files.

How to choose

  1. Find the smallest model that does your task. Test on the computer you own or on a rented hour. Which models suit which work is in best local LLMs by hardware and task.
  2. Count the people who will ask at the same moment, not the people in the company. Each one is a slot, and each slot at 8k costs 1.07 GB on the 8B, 1.61 GB on the 14B, 1.34 GB on the 24B and 2.68 GB on the 70B.
  3. Do the sum and find the first row of the fit table that says “yes” for your case.
  4. Then look at the price, and at who will look after the machine.

On the second point we have numbers. Ten simultaneous requests to a 7B model on a Spark each got a full answer in about 15 seconds with vLLM (receipt). With Ollama as installed, one slot, ten requests on our other Spark waited a median of 71.7 seconds for the first word (receipt). Same hardware, different setup. The office setup, with accounts and logs, is in our Open WebUI office guide, and the levels of “private” in a private ChatGPT alternative for your business.

Worked through for three common cases:

  • One person, drafts and summaries, 8B. Any 12 GB card, or the PC you own if you accept waiting.
  • A team of ten, 8B to 14B, four asking at once. The 8B fits a 12 GB card; the 14B needs 24 GB. Among complete machines, a 24 GB Mac mini holds either.
  • A 32B or 70B. First check that the 14B or 24B really fails your task. Then 32 GB for the 32B at a short context, and 64 GB of unified memory for the 70B.

The second-hand card route

A used 24 GB card is still the cheapest way to 24 GB: EUR 900 to 1,400 asked, against EUR 2,058.49 for the cheapest new card with more. The risks are real.

  • The guarantee depends on who sells. This is information, not legal advice. Spanish consumer law makes a business answer for faults for three years, and lets it agree a shorter period for second-hand goods, never under one year (Article 120.1). It covers sales by a business to a consumer (Article 2), so a private seller on a classifieds app is outside it. Ask before you pay.
  • You cannot see its history. Ask for a stress test on video, and run your own the day it arrives.
  • Power and space. NVIDIA asks for a 750 W supply for one RTX 3090. Two cards need the supply, the slots and the airflow for 700 W of cards alone.

Power and noise

The makers’ figures are maximums, not what a machine draws while it waits: 155 W for a Mac mini, 480 W for a Mac Studio, a 240 W supply on a DGX Spark, 7 to 25 W for a Jetson. A tower adds the card to the rest of the PC: 180 W for an RTX 5060 Ti, 575 W for an RTX 5090.

To turn watts into money use the formula in the pillar guide: watts × hours per day × 30 ÷ 1000 × your price per kWh. Measure at the plug if you can. We have not measured the draw of any machine in this catalogue.

On noise, Apple publishes 5 dBA for the Mac mini at idle. No other maker in the table gives a figure on the pages we read.

What to avoid

  • Buying for the largest model you could load. A 70B at 4.42 tokens per second takes almost two minutes for a 500-token answer.
  • An 8 GB card or an 8 GB board for a team. An 8B needs 6.49 GB for one person and 9.71 GB for four at once.
  • A Raspberry Pi 5 as the AI machine. Its 16 GB model is USD 305 on raspberrypi.com. A 12 GB card in a PC you own is a better use of that money.
  • A marketplace seller’s price read as the shop’s. Two near-identical PNY RTX 5060 Ti 16 GB cards were EUR 859.99 from the retailer and EUR 1,321.90 from a third-party seller on the same site.
  • Leaving the context at its default. It can multiply the memory the model takes: see Ollama num_ctx memory usage, measured.

Frequent questions

Is 16 GB enough? For an 8B at any context here, yes. For a 14B at 8k, yes on a card and tight on a Mac. Not for a 24B.

Mac or PC with a card? A card is cheaper up to 16 GB if you own the PC. Above 32 GB this table has no card, only unified-memory machines. We have measured neither, so we will not say which is faster.

Can I add memory later? Not on a Mac, a DGX Spark or a Framework Desktop: the memory is fixed at purchase. On a PC you can change the card.

Is the Mac mini M4 still a good buy? Apple’s Spanish store no longer lists it; the mini starts with the M6. We could not source a current new price, so it is not in the table.

Limits

  • Prices are one day, one country and mostly one retailer. The second-hand range is asking prices from eleven listings, not sales.
  • We own two DGX Sparks and nothing else here. Every other speed is someone else’s or absent.
  • The fit table is a planning sum, checked against loads on our own machines only.
  • We did not measure power, noise or heat, and we did not price laptops, AMD cards or rented GPUs.

Next steps

Let’s talk for 15 minutes

We size hardware from your real documents and the number of people asking at once, and we say so when the computer you own is enough. If you want a second opinion before you order, ask for a 15-minute call or see how we work in consulting.

Diagram
Share: LinkedIn X
Veredicto semanal

Get new guides before anyone else

Subscribe and we tell you when new guides, templates and workflows go up. One email a week, no spam.

Already published: 102 guides and 25 templates. All free, no signup.

Bonus: the local-AI starter pack PDF when you subscribe
Once a week No spam Unsubscribe anytime

See what you get