Local AI: what hardware for what budget?
Updated on 23 October 2026. Prices change every week: we revise this article regularly (update log at the end of the page).
We have explained why an AI installed on your premises (in French) makes sense when your documents are confidential. This article answers the next question, the one our clients always ask: which machine should you buy, and what will you get?
We selected six configurations available from a Belgian retailer, Coolblue, and matched each with the open AI model that gets the most out of it. The speeds quoted come from published measurements (specialist forums, benchmarks, test videos); where only an estimate exists, we say so.
Why hardware is so expensive in 2026
If you have bought a computer this year, you have noticed: anything with a lot of memory has soared.
- Memory makers (Samsung, SK Hynix, Micron) are reserving their output for AI data centres, which are far more profitable. RAM prices have more than doubled since late 2025, with a record rise of 90 to 95% in the first quarter of 2026 alone (TrendForce).
- The pressure is not easing: TrendForce still expects a 10 to 15% rise in the last quarter of 2026, and Micron says 2027 will be tighter than 2026. Micron has even dropped its consumer brand, Crucial.
- Apple has raised its prices: the entry-level Mac mini went from €719 (2024) to €1,069 (2026).
- The most powerful consumer graphics card, the RTX 5090, sells for between €2,789 and €5,259 at Coolblue Belgium, and every model is out of stock there. The only complete PC that includes one costs €7,999.
The French channel Underscore_ devoted an episode to this shortage, with a guest who runs 600 graphics cards in his company. He explains why factories switched to memory for AI chips and expects the shortage to last until 2028. He also claims that this memory accounts for "100%" of production: that is an exaggeration; TrendForce puts it at about half of 2026 volumes, including server memory.
What this means for you: do not wait for a quick price drop, and be wary of 2025 comparisons, whose prices no longer exist.
The only figure that matters: memory
An AI model must fit entirely in fast memory to respond quickly. Two families of machines compete.
- PC with an NVIDIA graphics card. The model has to fit in the card's memory (the "VRAM"), which is very fast but limited to 16 GB on affordable cards. If it spills over into the PC's main memory, speed collapses. Result: very fast answers, but medium-sized models.
- Apple Mac. The processor and the graphics part share the same memory ("unified memory"). You can therefore load much larger models, at a lower speed than an NVIDIA card of the same size. In return: silence, low power consumption and no hardware tuning.
Maxime, chief technology officer of the start-up Gladia, summed it up on Underscore_ in September 2026: until recently, local models were "at least two years behind"; recent models from Qwen and Google (Gemma) and the Mac's unified memory have changed the game, to the point that he is moving his own assistant to a Mac mini. He also notes that for reliable French text, Mistral is often still the best choice among open models.
A recent development changes the picture: so-called "mixture of experts" (MoE) models. A 120-billion-parameter model only uses about 5 billion for each piece of text it produces: you need a lot of memory to load it, but it answers quickly. This is what makes Macs with large memory interesting.
How to read the speeds in this article: we give them in words per second (strictly speaking "tokens", pieces of words). An adult reads silently at about 4 words per second (around 240 words per minute in English, according to a review of 190 studies). Above 20, the answer appears faster than you can read it; below 10, you wait.
The six configurations at a glance
| # | Machine (Coolblue.be, 23/10/2026) | Price | Memory for AI | Recommended model | Speed |
|---|---|---|---|---|---|
| 1 | Mac mini M6 16 GB / 256 GB | €1,069 | ~12 GB usable | Qwen3.5 9B | ~27 words/s (measured) |
| 2 | Mac mini M6 24 GB / 512 GB | €1,509 | ~20 GB usable | gpt-oss-20b | no published measurement |
| 3 | Cobalt x MSI PC, RTX 5060 Ti 16 GB, 32 GB RAM | €1,759 | 16 GB (card) | gpt-oss-20b | 82 to 95 words/s (measured) |
| 4 | Mac mini M5 Pro 24 GB / 512 GB | €2,019 | ~20 GB, 1.8× faster than the M6 | Gemma 4 26B | no published measurement |
| 5 | PCSpecialist PC, RTX 5070 Ti 16 GB, 32 GB RAM — or Mac Studio M5 Max 36 GB | €2,949 / €3,029 | 16 GB very fast / ~32 GB | Qwen3.6 35B | see details |
| 6 | Mac Studio M5 Ultra 96 GB / 1 TB | €6,649 | ~90 GB | gpt-oss-120b | no independent measurement yet |
All the recommended models are published under the Apache 2.0 licence: free commercial use, no royalties.
1. Around €1,000: Mac mini M6 16 GB
For whom: a self-employed person who wants to summarise, rephrase and translate their own documents without sending anything online.
The model: Qwen3.5 9B (Alibaba, March 2026), about 6 GB in compressed form. A model of this size runs at about 27 words per second on a Mac mini M6, according to a measurement published by MindStudio: comfortable.
The limits: 16 GB, part of which goes to macOS and your applications. No room for a larger model, and a Mac's memory can never be expanded after purchase. We do not recommend this configuration for a team.
2. Around €1,500: Mac mini M6 24 GB
For whom: the workstation of someone who uses AI several hours a day.
The model: gpt-oss-20b (OpenAI, August 2025), about 13 GB. It is a mixture-of-experts model: it only activates about 3.6 billion parameters per word, which makes up for the M6's modest memory bandwidth (170 GB/s according to Apple). For an AI that also reads images (scanned invoices, screenshots), Google's Gemma 4 26B fits in 24 GB, with a reduced text window.
The limits: no independent measurement has been published for this machine and model combination so far. "Dense" 27-billion-parameter models are slow on it: 8 to 10 words per second measured on a 32 GB Mac mini M6.
3. Around €1,800: PC with an RTX 5060 Ti 16 GB
For whom: the best speed-to-price ratio in this comparison, for a small team sharing an AI over the network.
The model: gpt-oss-20b, which fits entirely in the card's 16 GB and answers at 82 to 95 words per second, even with long documents (Hardware Corner and ComputingForGeeks measurements). Qwen3.6 35B, more capable, slightly overflows the card and runs at around 44 words per second for its direct predecessor (InsiderLLM).
Watch out for the trap: at Coolblue, two PCs listed as "RTX 5060 Ti" at €1,799 (PCSpecialist Impact 20 and Cobalt "RTX 5060 Ti 8GB") only have 8 GB of graphics memory. For AI, that is half the space under the same name. Check for "16 GB" on the product page before ordering.
The limits: the noise and power consumption of a gaming PC (180 W for the card alone), and Windows to maintain.
4. Around €2,000: Mac mini M5 Pro 24 GB
For whom: anyone who wants the silence and frugality of a Mac, with a speed that gets closer to a PC.
The model: Gemma 4 26B (Google, April 2026), a mixture-of-experts model that reads text and images and was trained on more than 140 languages. The M5 Pro chip offers 307 GB/s of memory bandwidth, 1.8 times more than the M6: same models, noticeably faster answers.
The limits: Coolblue only sells the 24 GB version. The 48 and 64 GB versions, far more interesting for AI, can only be custom-ordered from Apple.
5. Around €3,000: RTX 5070 Ti PC or Mac Studio M5 Max
Two philosophies at the same price.
The RTX 5070 Ti PC (€2,949) keeps 16 GB on the card, but much faster: gpt-oss-20b exceeds 140 words per second. With a setting that keeps part of the model in main memory, Qwen3.6 35B reaches 43 words per second as standard and more than 110 with an optimised version of llama.cpp (measurements published on an RTX 5070 Ti). It handles simultaneous questions well: four parallel requests add up to about 380 words per second on gpt-oss-20b. It is a machine for a team.
The Mac Studio M5 Max 36 GB (€3,029) loads larger models, such as Qwen3.8 27B (Alibaba, August 2026), at the top of several October 2026 rankings for its size, which reads text, images and video. The downside: a "dense" model of this size tops out at around 25 to 30 words per second on a 40-core M5 Max chip (community measurements); the 36 GB, 32-core version has 25% less bandwidth. Qwen3.6 35B, a mixture-of-experts model, is much faster on it.
Our view: the PC for speed and teamwork, the Mac for answer quality, silence and power consumption.
6. Around €6,600: Mac Studio M5 Ultra 96 GB
For whom: an SME that wants the most capable open model that fits on a single desktop machine, for the whole team, without any document leaving the building.
The model: gpt-oss-120b (OpenAI), about 65 GB, which fits in 96 GB with room to spare for long documents. On a Mac with a 128 GB M5 Max chip, it produces 65 to 88 words per second (community measurements). The M5 Ultra chip has twice the bandwidth (1.2 TB/s): it should do at least as well, but no independent measurement has been published yet.
The limits: €649 over the stated budget, and a machine to reserve for shared, daily use if it is to pay for itself.
What we left out
- PC with an RTX 5090 (32 GB on the card): the fastest for a medium-sized model (Qwen3.8 27B at 78 words per second, more than 110 with recent optimisations), but €7,999 at Coolblue.
- Mini PCs with an AMD Ryzen AI Max+ 395 processor and 128 GB (Framework, GMKtec, Minisforum, HP Z2 Mini G1a): a lot of memory for the money (about $3,300 to $4,300 in September 2026), gpt-oss-120b at 34 to 38 words per second (StarryHope). Not sold at Coolblue, and their price has doubled since June.
- NVIDIA DGX Spark: designed for developers, slow at generating text for its price.
- Second-hand graphics card (RTX 3090): the enthusiasts' favourite, but without a warranty.
- Mac Studio M5 Max 128 GB: excellent, but only available made to order from Apple.
- Renting a graphics card in a European data centre: an alternative without purchase, which the YouTuber xavki demonstrates with an NVIDIA L4 card at the Swiss host Infomaniak. Relevant for occasional use.
Screen, keyboard, backup: the additional budget
Our prices are for the machine alone. Also allow for:
- A screen: €150 to €400 for a standard office model. Unnecessary if the machine only serves as the team's AI server.
- Keyboard and mouse: not supplied with a Mac mini or a Mac Studio.
- An uninterruptible power supply (UPS) if the machine serves the whole team: a power cut should not shut it down abruptly.
- A backup of the settings and indexed documents, following the 3-2-1 rule.
- Electricity: a Mac uses little power; a PC with a graphics card, left on permanently, weighs on the bill. Configure standby.
Three ways to go about it
| Option | Cost | Time | Risk |
|---|---|---|---|
| Do nothing: stay on ChatGPT, Copilot, Gemini | $20 to $200 per month per user, and rising | None | Your documents leave the company; dependence on one supplier |
| Do it yourself | The machine | One to two days to choose, install and test | Wrong model, machine exposed on the network, no backup |
| Get support | The machine and a few hours of service | Half a day on your side | The lowest: defined use, secured network, trained team |
The paradox is well explained by the French channel Grand Angle Eco: the price of an AI answer is collapsing year after year, but the big providers' subscriptions keep rising. A machine bought once steps outside that logic.
Example: a six-person accounting firm
Illustrative case. A firm wants to summarise letters, prepare draft replies and query its internal procedures without sending client data online.
We would start with the RTX 5060 Ti 16 GB PC (€1,759) and gpt-oss-20b: fast enough for six people who do not all query it at the same time, and able to read long documents. If the team needs to read scans or reply in Dutch with more nuance, the Mac Studio M5 Max (€3,029) and Qwen3.8 27B would take over. In both cases: access limited to the internal network, individual accounts, backup.
Test yourself
At ExsIT
We use a local AI ourselves every day. The first contact is free: we listen to your needs, then propose an action plan. Choice of machine according to your actual use, installation, securing access, team training.
For your IT person
The details (speed estimation, quantisation, settings, security) are in our technical guide to local AI. In summary:
Software. Ollama or LM Studio to get started; llama.cpp (server compatible with the OpenAI API) for fine-tuning; vLLM to serve several simultaneous users on an NVIDIA card; MLX on a Mac.
Model sizes in 4 bits (Q4, MXFP4).
| Model | Type | Active parameters | Memory |
|---|---|---|---|
| Qwen3.5 9B | dense | 9 B | ~6 GB |
| gpt-oss-20b | MoE | ~3.6 B | ~13 GB |
| Gemma 4 26B-A4B | MoE, vision | ~3.8 B | ~15-18 GB |
| Qwen3.8 27B | dense, vision, 262K context | 27 B | ~17 GB |
| Qwen3.6 35B-A3B | MoE | 3 B | ~22 GB |
| gpt-oss-120b | MoE | ~5.1 B | ~63-66 GB |
Published measurements (generation, single user).
| Hardware | Model | Speed | Source |
|---|---|---|---|
| Mac mini M6 | dense 9B Q4, llama.cpp | 26.9 tok/s | MindStudio |
| Mac mini M6 32 GB | Qwen 27B 4-bit, MLX | 8-9 tok/s; 16-19 with multi-token prediction | StarryHope (YouTube tests cited) |
| RTX 5060 Ti 16 GB | gpt-oss-20b MXFP4 | 82-95 tok/s | Hardware Corner, ComputingForGeeks |
| RTX 5060 Ti 16 GB | Qwen3.5 35B-A3B Q4, 100K context | 44 tok/s | InsiderLLM |
| RTX 5070 Ti 16 GB | gpt-oss-20b; Qwen3.6 35B-A3B Q4 | 144; 43 (standard) to 112 (expert cache) tok/s | GenerelSchwerz/llama.cpp |
| RTX 5090 32 GB | Qwen3.8 27B Q4_K_M | 78 tok/s; 118 with MTP | inferris |
| M5 Max 128 GB | gpt-oss-120b Q8, MLX | 88 (4K) / 65 (32K) tok/s | Hardware Corner |
| M5 Max 128 GB | Qwen3.6 35B-A3B Q8, llama.cpp | ~100 tok/s | stared/benching-local-llms |
Useful settings. 8-bit KV cache to extend the context with no measurable loss; on a 16 GB card, keep part of an MoE model's experts in main memory (--n-cpu-moe); multi-token prediction (MTP) on Qwen3.8 27B, about ×1.8 with no loss since every token is verified.
Security. Hundreds of misconfigured Ollama servers were found exposed on the Internet in 2026 (see the Wikipedia article on Ollama and its references). Listen on the local address or internal network only, never on the public address; authentication in front of the API; no port forwarding.
Update log
- 23 October 2026: first version. Prices checked on coolblue.be.
Sources
Apple. (n.d.). Mac Studio – Technical specifications. https://www.apple.com/mac-studio/specs/
AppleInsider. (2026, 29 September). M5 Max Mac Studio review: Far better choice than Ultra for most desks. https://appleinsider.com/articles/26/09/29/m5-max-mac-studio-review-far-better-choice-than-ultra-for-most-desks
Brysbaert, M. (2019). How many words do we read per minute? A review and meta-analysis of reading rate. Journal of Memory and Language, 109, 104047. https://doi.org/10.1016/j.jml.2019.104047
ComputingForGeeks. (2026). Ollama models cheat sheet: Best models by VRAM. https://computingforgeeks.com/ollama-models-cheat-sheet/
ComputingForGeeks. (2026). RAM prices in 2027. https://computingforgeeks.com/ram-prices-2027/
Coolblue. (2026). Mac mini, Mac Studio, gaming PCs RTX 5060, RTX 5070 and 5070 Ti, RTX 5090 [product pages consulted on 23 October 2026]. https://www.coolblue.be/fr/ordinateurs/apple/mac-mini
DataNorth. (2026, 14 August). Alibaba releases Qwen3.8-27B. https://datanorth.ai/news/alibaba-releases-qwen3-8-27b
DataNorth. (2026, 3 April). Google releases Gemma 4: Four open models with Apache 2.0 license. https://datanorth.ai/news/google-releases-gemma-4-open-models
GenerelSchwerz. (2026). Benchmark comparison showcase [llama.cpp, RTX 5070 Ti]. GitHub. https://github.com/GenerelSchwerz/llama.cpp/wiki/Benchmark-Comparison-Showcase
Grand Angle Eco. (2026, 23 August). La vérité terrifiante derrière l'effondrement du prix de l'IA [Video]. YouTube. https://www.youtube.com/watch?v=-K0E6yNhgpw
Hardware Corner. (2026). Apple M5 Max for local LLMs: First benchmarks vs RTX Pro 6000 and RTX 5090. https://www.hardware-corner.net/m5-max-local-llm-benchmarks-20261233/
Hardware Corner. (2026). RTX 5060 Ti 16GB local LLM benchmarks. https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/
iCulture. (2026). Mac mini 2026 prijzen. https://www.iculture.nl/nieuws/mac-mini-2026-prijzen/
inferris. (2026). inferris [Qwen3.8-27B, RTX 5090]. PyPI. https://pypi.org/project/inferris/
InsiderLLM. (2026). RTX 5060 Ti review for local AI. https://insiderllm.com/guides/rtx-5060-ti-local-ai-benchmarks/
MacRumors. (2026). 2025 vs. 2026 Mac Studio buyer's guide. https://www.macrumors.com/guide/2025-vs-2026-mac-studio/
MindStudio. (2026). M6 Mac mini local AI performance. https://www.mindstudio.ai/blog/m6-mac-mini-local-ai-performance
PCWorld. (2025). DRAM, SSD shortages could last months to years, vendors say. https://www.pcworld.com/article/3028843/dram-ssd-shortages-could-last-months-to-years-vendors-say.html
stared. (2026). Benching local LLMs on Apple Silicon [M5 Max 128 GB]. GitHub. https://github.com/stared/benching-local-llms-on-apple-silicon
StarryHope. (2026). Mac mini M6 vs M5 Pro for local AI. https://www.starryhope.com/minipcs/mac-mini-m6-vs-m5-pro-local-ai/
StarryHope. (2026). The DGX Spark and the reality of local AI hardware in 2026. https://www.starryhope.com/ai/dgx-spark-local-ai-hardware-landscape/
The Decoder. (2026). Alibaba's Qwen team releases Qwen 3.8 models with open weights under the Apache 2.0 license. https://the-decoder.com/alibabas-qwen-team-releases-qwen-3-8-models-with-open-weights-under-the-apache-2-0-license/
Underscore_. (2026, 26 May). Cette pénurie de RAM arrange beaucoup trop de monde [Video]. YouTube. https://www.youtube.com/watch?v=gtIwKyWDtj8
Underscore_. (2026, 29 September). Il a laissé l'IA gérer sa vie pendant 6 mois [Video]. YouTube. https://www.youtube.com/watch?v=SeQXCpZhbjc
Wikipedia. (2026). Ollama [consulted on 23 October 2026]. https://en.wikipedia.org/wiki/Ollama
xavki. (2026, 22 September). vLLM - 05. Premier inférence (model Mistral) [Video]. YouTube. https://www.youtube.com/watch?v=F5_7SYX8C38