Local AI Hardware Catalogue 2026: What to Buy at EUR 0, 250, 700, 1,000 and 2,500
By the end of this post you will know which budget tier your business belongs in, what that money buys in September 2026, and which model each device runs at a usable speed. The cheapest tier costs nothing: it is the computer already on your desk. Our Edge AI hardware guide compares three devices in depth; this catalogue is wider and organised by what you can spend.
What you need
- Ten minutes and the five questions below.
- The rough size of the model your task needs (3B, 8B, 14B, 32B or 70B parameters). If you do not know yet, start at 8B: it is the size most business tasks settle on.
- Any AI assistant to check your arithmetic. Paste the tables in and ask: “Given my model size and number of users, which tier do I need?”
Prices are what retailers and trackers published in 2026, converted loosely; memory prices moved a lot this year, so check the day you buy. Speeds are tokens per second (tok/s) on 4-bit quantized models. Every number we did not measure ourselves links to its source.
Step 1: Answer five questions before you spend anything
- What is the largest model your task actually needs? Test it first on the machine you have or on a free hosted demo. Most document, email and support tasks work at 8B. Reasoning over long contracts wants 14B to 32B. Only a few tasks justify 70B.
- How many people will use it at the same moment? One person at a time is a mini PC job. Five concurrent users need a discrete GPU, because throughput scales with memory bandwidth.
- Where will the machine sit? A silent box on a desk points to a Mac mini or a Jetson. A noisy tower in a cupboard opens up second-hand GPUs, which give the most memory per euro.
- Do you already own a PC with a free PCIe slot and a 550 W power supply? If yes, the cheapest upgrade is a graphics card, not a new computer.
- Who will maintain it? NVIDIA and Apple have the smoothest software path. Intel Arc gives the cheapest video memory but needs more setup, as we cover in the limits section.
Write down your answers. The tier table below is read from them.
Step 2: Read the catalogue by tier
| Tier | Device | Memory for models | Runs comfortably | Speed (third-party unless noted) |
|---|---|---|---|---|
| EUR 0 | The laptop or PC you own (16 GB RAM, CPU only) | shared system RAM | 3B; 8B slowly | see our cost post for how to measure it |
| EUR 250 | Raspberry Pi 5, 16 GB | 16 GB shared | 1B to 3B | 3.3 to 5.8 tok/s on Llama 3.2 3B (aidatatools) |
| EUR 250 | NVIDIA Jetson Orin Nano Super | 8 GB shared | 3B | 12 to 18 tok/s on 3B (Edge AI Vision) |
| EUR 250 (card only) | Used RTX 3060 12 GB, in a PC you own | 12 GB VRAM | 8B, 14B tight | 52 to 65 tok/s on Llama 3.1 8B Q4_K_M (SpecPicks); price $239 to $280 (GPUDojo) |
| EUR 250 (card only) | Intel Arc B580 12 GB, in a PC you own | 12 GB VRAM | 8B | 73.8 tok/s on Llama 2 7B Q4_0 with llama.cpp SYCL (llama.cpp discussion); $249 new (InsiderLLM) |
| EUR 700 | Mac mini M4, 16 GB | 16 GB unified | 8B; 14B tight | EUR 719 (Applesfera); same chip as the 24 GB row |
| EUR 700 (card only) | RTX 5060 Ti 16 GB, in a PC you own | 16 GB VRAM | 8B to 14B | EUR 550 to 700 (VideoCardz) |
| EUR 1,000 | Mac mini M4, 24 GB | 24 GB unified | 8B to 14B; 32B tight | ~35 tok/s on Qwen 2.5 7B (Compute Market); EUR 719 + 250 for 24 GB (Applesfera) |
| EUR 1,000 (card only) | Used RTX 3090 24 GB, in a PC you own | 24 GB VRAM | 14B to 32B | $800 to $1,200 (XDA); 70B only with CPU offload at 5.2 tok/s (GigaGPU) |
| EUR 2,500 | Mac Studio M4 Max, 36 GB | 36 GB unified | 32B | EUR 2,329 (Applesfera) |
| EUR 2,500 | Tower with two used RTX 3090 (48 GB) | 48 GB VRAM | 70B at 4-bit | two cards alone are $1,600 to $2,400; a pair of RTX 4090 reaches 21 tok/s on Llama 3 70B (Sergey Nog), 3090s are slower |
Two rows are deliberately just above the tier. A MacBook Pro M4 Max with 64 GB runs Llama 3.3 70B at 12.5 to 17.2 tok/s (Markaicode), and the new Mac Studio M5 Max starts at EUR 3,029 (Hipertextual). Comfortable 70B starts above EUR 2,500. Below it, 70B means offloading to CPU and single-digit speeds.
On the Raspberry Pi price: it launched at $120 (Raspberry Pi), rose by $60 in February 2026 because of memory costs (Raspberry Pi), and Tom’s Hardware later reported $205 (Tom’s Hardware). At that price the Jetson is the better buy for AI.
Step 3: Check the model fits in memory
A 4-bit model file needs roughly its file size plus room for context. Use this rule of thumb, then confirm with ollama ps after the first run:
# file sizes at Q4_K_M, from the Ollama library and Hugging Face
# 3B ~2 GB | 8B ~5 GB | 14B ~9 GB | 32B ~20 GB | 70B ~43 GB
ollama run llama3.1:8b "Say hello in one line"
ollama ps # SIZE column = what the model really occupies right nowLeave 20 percent free. If ollama ps shows a split between CPU and GPU, the model does not fit and you are in the tier below the one you thought.
What we measured
For scale, and not as a recommendation, here is our own workstation next to the tiers. It is an NVIDIA GB10 with 128 GB of unified memory, shared with about fifty other services, measured on 2026-09-09 with Ollama 0.30.10.
| Machine | Model | tok/s | Resident memory | Date |
|---|---|---|---|---|
| GB10, 128 GB | llama3.1:8b (Q4_K_M) | 23 to 24 (37.8 the day before, on a quieter box) | 5.3 GB at 4k context | 2026-09-09 |
| GB10, 128 GB | qwen3.6:35b (Q4_K_M) | 37 to 60 across three runs | 23 GB at 8k context | 2026-09-09 |
| GB10, 128 GB | qwen2.5vl:72b (Q4_K_M) | 3.2 (cold run) | 8k context | 2026-09-09 |
The lesson is in the last row: even 128 GB of unified memory gives single-digit speeds on a 72B model, because generation speed follows memory bandwidth, not memory size. A used RTX 3060 beats our workstation on an 8B model. Buy for the model you will run, not for the biggest one you could load.
Where this fits, and the honest limits
- EUR 0 to 250 covers a single-user assistant on 3B to 8B models: drafting, classification, summaries of short documents. It will not handle five people at once.
- EUR 700 to 1,000 is where most small businesses land: an 8B to 14B model serving a team of ten for routine work, silently, at under 40 W.
- EUR 2,500 buys 32B comfortably or 70B slowly. Before spending it, check that the 14B model really fails your task; often it does not.
- Intel Arc is the cheapest video memory in the table, but Ollama has no native Arc support and the llama.cpp backend needs setup (InsiderLLM). Choose it if someone on your team enjoys Linux.
- Second-hand cards carry no warranty. Buy from a seller who shows a stress test, and budget a power supply.
Next steps
- Read the three-device comparison first if you have not: Edge AI hardware guide 2026.
- Work out the month a device pays for itself: Cloud vs local break-even in 15 minutes.
- Fit a bigger model into the memory you have: Quantization explained.
- Install the software once the box arrives: How to deploy AI locally in your business.
Work with us
We size hardware for clients from their real token counts and concurrency, and we are happy to tell you that the EUR 0 tier is enough when it is. If you want a second opinion before you order, get in touch or read how our consulting works.