Local AI on your Mac
Use AI without your business data ever leaving your Mac. What local AI is, which Mac and tools you need, what it costs and when the cloud is still the better choice.
AI is now part of everyday work at most SMEs. According to the Exact SME Barometer 2026, a survey of more than 1,300 Belgian SMEs with 2 to 250 employees, 81% of SMEs use AI, up from 8% in 2024. Exact At the same time, half say they don't really know how to put AI to work. Exact And every question put to a cloud AI service raises the same issue: where does the information your employee types in actually end up?
Local AI is one answer. The AI model runs on your own Mac, and your texts, customer data and documents stay on that device. This guide explains what local AI is, which Mac you need, which free tools and models are available, what it costs and when a cloud service is still the better choice.
In short
What is local AI?
With ChatGPT, Copilot or Claude, every question travels over the internet to the AI company's servers. That is where the model runs and where the answer is produced. With local AI, the model runs on your own computer. Your Mac does the processing, and whatever you type never leaves the device.
Think of it this way: a cloud service is a translator you send your documents to. They do excellent work, but your documents pass through their office. Local AI is a translator who sits in your own office. Everything stays in-house.
Your data stays at home
Nothing is passed to an AI company, so there is no debate about where that server is.
Works offline
After a one-off download of a model, you no longer need an internet connection.
No subscription
The tools and models are free. You invest in the Mac, not in a monthly licence per user.
Local AI doesn't replace everything ChatGPT or Copilot can do. The models that fit on a Mac are smaller than the largest models in the cloud. For an SME's everyday reading and writing, such as drafting emails, summarising reports, translating texts or rewriting a quote, they are perfectly usable. Where the limits lie is covered further down under the limitations.
Why local: your data, the GDPR and the cloud
When you use AI in the cloud, you pass data to another company. That isn't forbidden, but you need to know what you are passing on and on what terms. In practice, that often goes wrong.
- Staff type in sensitive information. According to Cyberhaven's AI Adoption & Risk Report 2025, 34.8% of the business data employees enter into AI tools is sensitive, up from 10.7% in 2023. Cyberhaven
- Businesses have little oversight. In a 2025 BigID survey, 69% of organisations name AI-related data leaks as their biggest concern, and 47% have no security measures aimed specifically at AI. BigID Metomic
- Free and personal accounts learn from your conversations. On ChatGPT Free, Plus and Pro, OpenAI uses your conversations to improve its models by default, unless you switch this off under Data Controls. On the business plans and the API, that doesn't happen by default. OpenAI Help Center
- Even business services can process data outside Europe. Since April 2026, Microsoft 365 Copilot has had a setting for European customers called flex routing. Office 365 for IT Pros When it is on, Microsoft may carry out the processing in the United States, Canada or Australia at times of peak demand. In new environments it is on by default. Stored data stays in the EU, apart from limited pseudonymised data, and an administrator can switch flex routing off in the Microsoft 365 admin centre. Microsoft Learn
For the GDPR, this means: if you work with personal data about customers or staff, you need a data processing agreement with the AI provider, and any transfer outside the EU belongs in your risk assessment. If the risk is high, the GDPR requires a data protection impact assessment (DPIA). Breaching the basic principles can lead to a fine of up to 20 million euros or 4% of worldwide annual turnover. EUR-Lex
Regulators are watching. In December 2024, the Italian data protection authority fined OpenAI 15 million euros. The court in Rome overturned that decision in March 2026 Garante Privacy, but the case does show what European regulators pay attention to.
Why local makes it so simple. If the model runs on your own Mac, you pass nothing to an AI provider. No data processing agreement with an AI company is needed for that processing, and the question of whether a server is in Europe disappears. You are still responsible for the Mac itself, though: encryption, updates and who has access to it.
Belgium now uses AI more than most European countries. According to Eurostat, 34.5% of Belgian businesses with at least ten employees used at least one AI technology in 2025, compared with an EU average of 20%. That puts Belgium fourth in the EU, after Denmark, Finland and Sweden. Eurostat
Which Mac do you need?
Any Mac with Apple silicon, meaning an M1 chip or newer, can run local AI. Two things determine what you can do with it:
- How much memory it has. A Mac with Apple silicon has unified memory: the processor and the graphics chip share the same memory. The AI model has to fit into it completely. The more memory, the larger and smarter the model you can load.
- How fast that memory is. How quickly the answer appears on your screen depends heavily on memory bandwidth. A standard M3 chip reaches 100 GB/s, an M5 Max up to 614 GB/s: six times as much. Apple Support Apple
How much memory for which model?
A rule of thumb: a model compressed to 4 bits takes up about 0.6 GB of memory per billion parameters. Google, for example, gives 6.7 GB for the 12-billion-parameter version of its Gemma 4 models and 17.5 GB for the 31-billion version. Google AI On top of that, you need memory for macOS, your other programs and the conversation itself. So allow plenty of headroom.
| Memory in your Mac | What fits comfortably | Suitable for |
|---|---|---|
| 8 GB | Small models, such as Gemma 4 E4B (4.5 GB) | Trying it out. For daily work it is tight. |
| 16 GB | Models up to around 12 billion parameters, such as Gemma 4 12B (6.7 GB) | Emails, summaries, translations and short documents for one user |
| 24 to 32 GB | Models up to around 30 billion parameters, such as Gemma 4 31B (17.5 GB) or gpt-oss-20b | Longer documents and more complex questions |
| 64 GB and more | Large models, such as gpt-oss-120b (65 GB download) | A shared AI server for a team |
Memory use at 4 bits, for the model alone, according to Google (Gemma 4) and Ollama (gpt-oss). Google AI OpenAI states that gpt-oss-20b runs within 16 GB of memory. gpt-oss model card
The Mac range for local AI in October 2026
| Mac | Chip | Memory | Bandwidth | Price from, incl. VAT |
|---|---|---|---|---|
| MacBook Air | M5 | 16 to 32 GB | 153 GB/s | €1,429 |
| Mac mini | M6 | 16 to 32 GB | 153 to 170 GB/s | €1,069 |
| Mac mini | M5 Pro | 24 to 64 GB | 307 GB/s | €2,019 |
| MacBook Pro | M5, M5 Pro or M5 Max | 16 to 128 GB | 153 to 614 GB/s | See apple.com |
| Mac Studio | M5 Max or M5 Ultra | 36 to 512 GB | 460 GB/s to 1.2 TB/s | €3,029 |
Specifications according to Apple, prices from the Belgian Apple Store on 8 October 2026, for the base configuration. More memory costs extra. Apple Apple Newsroom Apple Store
Sole trader or small team
A MacBook Air with M5 and 16 GB, or better still 24 GB. Each user runs their own model, on the move too.
A shared AI server
A Mac mini with M5 Pro and 48 or 64 GB, in the office. Colleagues use the model over the network.
Large models, many users
A Mac Studio with M5 Max or M5 Ultra, for the largest open models or a bigger team.
Choose the memory when you buy. On a Mac with Apple silicon, you can't add memory afterwards. If you are torn between two configurations, take the larger one. We advise against the MacBook Neo, with 8 GB that can't be upgraded, for local AI.
Need a Mac for your team, chosen to suit every role and paid for up front or in instalments? We arrange that through Apple for your business.
Which models can you run locally?
A model is the actual AI brain; the software is just the packaging around it. Several companies release their models for download, with a licence that allows business use. For an SME in Flanders, one extra question matters: how good is the model's Dutch?
| Model | From | Sizes | Licence | Dutch |
|---|---|---|---|---|
| Gemma 4 Google AI | E2B, E4B, 12B, 26B and 31B | Apache 2.0 | Yes, more than 140 languages | |
| Qwen 3.5 Qwen model card | Alibaba | From 0.8B to 122B | Apache 2.0 | Yes, 201 languages and dialects |
| Ministral 3 and Mistral Small 4 Mistral AI | Mistral AI, France | 3B, 8B and 14B; Small 4 is larger | Apache 2.0 | Yes, explicitly listed |
| gpt-oss gpt-oss model card | OpenAI | 20b and 120b | Apache 2.0 | Not listed |
| Phi-4 Phi-4 model card | Microsoft | 14B | MIT | Limited, trained mainly on English |
| Llama 4 Llama 4 model card | Meta | Scout (109B) and Maverick (400B) | Llama 4 licence | Not in the list of supported languages |
"12B" stands for 12 billion parameters. Languages and licences according to the makers, as of October 2026.
For Dutch texts, Gemma 4, Qwen 3.5 and the Mistral models are the logical first choice today. Meta's Llama 3.1, the model everyone recommended last year, dates from July 2024 and has since been overtaken. Llama 4, on the other hand, is too large for most Macs. Llama 4 model card
New models follow each other quickly. We update this table when a new generation comes out that makes a difference for SMEs.
The tools: Ollama, LM Studio and Open WebUI
To run a model, you need a program that loads it and puts a chat window around it. The three best known are free.
| Ollama | LM Studio | Open WebUI | |
|---|---|---|---|
| What it is | App and background service that runs models | App with a chat window and model library | Chat environment in the browser, on top of Ollama |
| Price | Free, open source (MIT) | Free, at work too LM Studio | Free, with its own licence |
| Mac | macOS 14 or later; Apple silicon, Intel on the processor only Ollama Docs | macOS 14 or later, Apple silicon only | Through the browser |
| Without Terminal | Yes, with the app since July 2025 Ollama Blog | Yes | Installation requires technical knowledge |
| For a team | Yes, through the API on the network | Possible, less common | Yes, with separate accounts for each user |
| Watch out | Models with "cloud" in the name run on Ollama's servers | Some extra features use cloud models | Above 50 users you have to keep the name and logo Open WebUI licence |
- Ollama is the engine we use most often. The app has a simple chat window where you can also drop PDF and text files. Ollama Blog On a Mac in the office, Ollama can act as a shared AI server, so colleagues can use the model without installing it themselves. By default, Ollama only listens to its own Mac; opening it up to the network is something you do deliberately and securely.
- LM Studio is the handiest choice if you want to compare models yourself. You search for a model, click download and start chatting. On Apple silicon it uses MLX, among other things, Apple's open-source framework for machine learning. MLX on GitHub Since July 2025, you no longer need a commercial licence to use it at work. LM Studio
- Open WebUI gives your team a familiar chat environment in the browser, with an account for each employee, while the processing happens on your own Mac.
Local isn't always local. Ollama recently started offering cloud models too, such as gemma4:cloud. These run on Ollama's servers, mostly in the United States. If you want to be sure everything stays on your Mac, don't choose a model with "cloud" in the name, or switch off the cloud features in Ollama's settings.
What about Apple Intelligence?
Your Mac also has AI built in. Apple Intelligence works on any Mac with M1 or newer and speaks Dutch. Apple Support The language model behind Apple Intelligence runs on the device itself, and apps can use it offline too through Apple's Foundation Models framework. Apple Newsroom For rewriting, summarising and organising in Apple's own apps, that is a useful foundation. If you want to choose which model you use, or want an AI server for your team, you need Ollama or LM Studio. Bear in mind, though: apps can also call cloud models through that same framework. Apple Newsroom
Get started in five steps
For a single user, a first version is quick to set up. This is how to go about it with Ollama:
- 1Check your Mac. In the Apple menu, choose About This Mac. It shows the chip (M1 or newer) and the memory. Ollama needs macOS Sonoma (14) or later. Ollama Docs
- 2Install Ollama. Download the app from ollama.com and drag it to your Applications folder, like any other Mac program. Ollama Docs
- 3Choose a model that fits your memory. With 16 GB, start with a model of around 12 billion parameters, such as Gemma 4 12B. The model is downloaded once. Don't choose a model with "cloud" in the name.
- 4Test it with real tasks. Have it rewrite an email, summarise a report or translate a text. Feel free to compare two models: the difference in Dutch is sometimes large.
- 5Set the ground rules. Agree with your team which information may go into the local AI, and which questions may go to a cloud service. Put that in writing: that is your AI policy.
At nxtbit we tested Ollama with Llama 3.1 on a MacBook Air M5 for an accountant. The model answered business questions about tax deadlines and drafted letters, completely offline, without a single word of client data leaving the device. The accountant had installed Ollama and got it running himself within 8 minutes.
Want it for a whole team, with a shared server, accounts for each employee and remote management? Then there is more involved: the network, security, model updates and agreements on use. That is the part we can set up and manage for you.
What does local AI cost?
With local AI, you mainly pay for the Mac. The software and the models are free. With AI in the cloud, you usually pay per user per month, or per use through an API.
| Local AI | AI in the cloud | |
|---|---|---|
| Software and models | Free | Subscription per user, or pay per use |
| Hardware | One-off: the Mac, with enough memory | No extra hardware |
| Costs as users are added | Unchanged, until the server reaches its limit | Rise with every user |
| Setup and maintenance | Installation, updates and security of the Mac | Settings, permissions and policy in the service |
A worked example. Microsoft 365 Copilot Business, the Copilot licence for businesses with up to 300 users, costs 21 dollars per user per month. Computerworld For five employees, that comes to 3,780 dollars over three years, excluding VAT. A Mac mini with M5 Pro, which can serve as a shared AI server, costs from €2,019 including VAT in Belgium.
That comparison isn't complete. Copilot works directly in Word, Excel, Outlook and Teams, and a local model can't do that. A local server, in turn, needs installing and maintaining. The real point is this: with local AI, the price doesn't go up with every extra user, and for confidential data you get a simpler privacy story.
The limitations of local AI
Local AI is strong for confidential reading and writing, but it also has clear limits:
Local or cloud: which question belongs where?
Most SMEs don't choose one or the other, but both. Local AI for everything confidential, a properly configured cloud service for the rest. Work through these questions from top to bottom. The first "yes" decides where your question belongs.
Does your question contain personal data about customers or staff, or a confidential document such as a contract?
Yes: localHave you agreed with a client that their data won't leave your organisation?
Yes: localDo you need up-to-date information from the internet, or do you want images generated?
Yes: cloudDo you want AI directly in Word, Outlook and Teams, working on documents already in Microsoft 365?
Yes: Copilot, properly configuredEither will do. Choose whatever works most simply for your team. In the cloud, though, always use a business account, never a free personal one.
The golden rule. Customer data, contracts and strategic documents don't belong in a free or personal AI account. Use local AI for them, or a business cloud service with a data processing agreement that your IT partner has configured properly.
How Copilot, Claude and local AI relate to each other, and which route suits your data, is also shown in the decision aid on our page AI for your business.
How nxtbit helps
We aren't tied to a single supplier: we work with Copilot, Claude and open-source AI. For local AI, we guide you from the first question to a working setup:
- 1Conversation and hardware. We look at which tasks you want to use AI for and assess your existing Macs free of charge during the introductory call.
- 2Choosing the model and tools. Together we choose the model that suits your tasks, your language and your hardware, and agree which data stays local.
- 3Installation and integration. Ollama, Open WebUI or LM Studio, a shared server in the office and the link with your existing software: we set it all up.
- 4Training and aftercare. Practical training for your team, without jargon, and optional maintenance through NXT managed services.
As a Jamf Professional Partner, we manage Apple devices for SMEs in Limburg, Antwerp, Flemish Brabant and East Flanders. Local AI fits seamlessly with that: the Mac running the model is encrypted, kept up to date and managed like every other device in your business.
Frequently asked questions about local AI on your Mac
Is local AI as good as ChatGPT?
Not for everything. For emails, summaries, translations and rewriting texts, local models are perfectly usable. For complex reasoning, up-to-date information and images, the big cloud services remain stronger. The advantage of local is that your data doesn't leave your Mac.
Can I enter customer data into ChatGPT or Microsoft Copilot?
Only in a business service with a data processing agreement, and if you have included the transfer in your risk assessment. With Copilot, flex routing can have the processing carried out outside the EU at times of peak demand, unless an administrator switches it off. Customer data doesn't belong in a free or personal account. With local AI, you don't pass it to an AI provider at all.
How much memory do I need for local AI on a Mac?
With 16 GB you can comfortably work with models of up to around 12 billion parameters. For larger models, choose 24 to 32 GB, and for a shared server for your team, 64 GB or more. You can't add memory afterwards, so be generous when you buy.
Does local AI also work on a Mac with an Intel processor?
Ollama works on a Mac with an Intel processor, but only on the processor, without the graphics chip. LM Studio doesn't work on an Intel Mac. For daily use, we recommend a Mac with Apple silicon, M1 or newer.
Can I use local AI without an internet connection?
Yes. After the one-off download of the tool and the model, everything works offline. That doesn't apply to the cloud models in Ollama: you can recognise those by "cloud" in the name.
Which model works best in Dutch?
Google's Gemma 4, Alibaba's Qwen 3.5 and the Mistral models explicitly support Dutch. Microsoft's Phi-4 was trained mainly on English, and Dutch isn't on Llama 4's list.
Is local AI really free?
The software and the models are: Ollama is open source and LM Studio is free to use at work. You pay for the Mac, with enough memory, and the time to set everything up and maintain it.
AI for your business
Copilot configured properly, Claude in your workflows or local AI on your Mac. With a policy that respects the GDPR.
Find out moreApple for your business
Macs to suit every role, delivered zero-touch and paid for up front or in instalments.
See the rangeWant local AI set up for you?
Together we choose the model and the Mac that suit your business, set everything up and teach your team how to use it. Already have a setup? Then we'll give it an honest review.
Sources
Exact, KMO Barometeronderzoek 2026 (SME Barometer, more than 1,300 Belgian SMEs, carried out by Markteffect in the first quarter of 2026), accessed 8 October 2026 via Exact.
Cyberhaven Labs, 2025 AI Adoption & Risk Report, April 2025, accessed 8 October 2026 via Cyberhaven.
BigID, AI Risk & Readiness in the Enterprise 2025, accessed 8 October 2026 via BigID, and cited by Metomic.
OpenAI, How your data is used to improve model performance, accessed 8 October 2026 via OpenAI Help Center.
Microsoft, Flex routing (EU and EFTA), updated 29 September 2026, accessed 8 October 2026 via Microsoft Learn.
Office 365 for IT Pros, Flex Routing for Microsoft 365 Copilot in Europe, 7 April 2026, accessed 8 October 2026 via Office 365 for IT Pros.
European Union, General Data Protection Regulation, Articles 28, 35 and 83, via EUR-Lex.
Garante per la protezione dei dati personali, decision no. 755 of 2 November 2024 and the notice on judgment no. 4153/2026 of the court in Rome (18 March 2026), accessed 8 October 2026 via Garante Privacy.
Eurostat, Use of artificial intelligence in enterprises, December 2025, and dataset isoc_eb_ai (updated 15 June 2026), accessed 8 October 2026 via Eurostat.
Apple, Apple unveils a more powerful Mac mini featuring the all-new M6 and M5 Pro, 25 August 2026, accessed 8 October 2026 via Apple Newsroom.
Apple, technical specifications for MacBook Air, MacBook Pro, Mac mini and Mac Studio, and MacBook Air (M3) Technical Specifications, accessed 8 October 2026 via Apple and Apple Support.
Apple Store Belgium, prices for MacBook Air, Mac mini and Mac Studio, accessed 8 October 2026 via Apple.
Apple, support article on the requirements and languages of Apple Intelligence, 14 September 2026, accessed 8 October 2026 via Apple Support.
Apple, Apple's Foundation Models framework unlocks new intelligent app experiences, September 2025, and Apple aids app development with new intelligence frameworks and advanced tools, June 2026, via Apple Newsroom and Apple Newsroom.
Google, Gemma 4 model overview (memory use) and Gemma 4 licence, accessed 8 October 2026 via Google AI for Developers.
Qwen, Qwen3.5 model card, accessed 8 October 2026 via Hugging Face.
Mistral AI, Ministral 3 and Mistral Small 4, accessed 8 October 2026 via Mistral AI.
OpenAI, gpt-oss-20b model card, accessed 8 October 2026 via Hugging Face.
Microsoft, Phi-4 model card, accessed 8 October 2026 via Hugging Face.
Meta, Llama 4 Scout model card, accessed 8 October 2026 via Hugging Face.
Ollama, documentation (macOS, FAQ, cloud), the new app (30 July 2025) and the model library, accessed 8 October 2026 via Ollama Docs and Ollama Blog.
LM Studio, LM Studio is free for use at work (8 July 2025) and System Requirements, accessed 8 October 2026 via LM Studio.
Open WebUI, licence, accessed 8 October 2026 via GitHub.
Apple Machine Learning Research, MLX, via GitHub.
Computerworld, Microsoft drops M365 Copilot price for SMBs, upgrades free Copilot Chat, 19 November 2025, accessed 8 October 2026 via Computerworld.