Your sensitive data leaves your hands
Customer records, finances, HR files and confidential documents end up on someone else’s servers.
Private AI platform
Private AI that runs in your office.
An AI assistant that works like ChatGPT, but runs on machines in your own office. Your team can ask questions, draft documents and search company files, and your data never leaves the building.
- Your company documents
- Your private AI assistantIn your office
- Answers that show their sources
In short
- Everything runs on Apple Mac Studio computers with M5 chips, in your office. Nothing is sent to outside AI companies.
- Your data stays under your control, which lowers the risk of breaking data protection rules.
- It uses open AI models (Qwen, Gemma and GLM) that are licensed for business use.
- You own the equipment. We look after it with a 60-day warranty and a monthly maintenance plan (minimum 12 months).
- Start small and grow: move from AI Box to AI Pro to AI Enterprise as more people use it. AI Pro and AI Enterprise work with your existing logins, and every package connects to your apps.
100%of your data stays in your office
385Wmaximum power for the largest setup (Apple spec)
8-10weeks from first meeting to launch
0data sent to outside AI services
The problem
Why sending company data to public AI is risky.
Popular AI tools send what you type to servers outside your company. If you handle customer, financial, HR or confidential information, that raises three problems.
It can break data rules
More and more privacy laws limit sending personal data overseas.
Costs keep climbing
You pay per person or per use, so the bill grows as more staff use it and is hard to budget.
The solution
What you get.
The AI brain
Proven open AI models that understand English and many other languages.
Answers from your documents
The assistant reads your own files before it answers, and shows where each answer came from.
An easy chat screen
A simple web chat for your team, with user accounts, access by department and a record of who asked what.
It all stays in your office
Every question is handled on your own machines. No cloud, no outside AI service.
How it works
What it runs on.
In your office
Small Apple computers
The system runs on an Apple Mac Studio with an M5 chip, sitting in your office. No cloud, no rented servers, nothing sent outside. It is a single quiet desktop: the largest setup draws at most 385W.
Grows with you
Three sizes
AI Box for a small team, AI Pro for several departments, and AI Enterprise on the M5 Ultra for the largest AI models.
Packages
Choose a size that fits your team.
We price each setup after a free assessment, based on how many people will use it, how many documents it needs to read, and what it needs to connect to. No commitment, no hidden fees.
AI Box
1 Mac Studio (M5 Max, 64GB)
For a small team trying private AI for the first time.
- For a small team or a single department
- Answers several questions at the same time
- Everything stays in your office
Technical: Qwen3.6-35B-A3B or Gemma 4 26B-A4B, document search, continuous batching
All-inclusive, no surprise costs. Includes:
- Machine setup, model installation, chat software and access control by department
- Loading all your internal documents within the assessed scope, including scanned documents that need OCR if identified in the assessment, and calibrating answer quality. The AI looks up your documents (RAG); the model is not retrained
- Testing, go-live and staff training
- 60-day warranty, then monthly maintenance and loading of new documents under the stated monthly fee
AI Pro
1 Mac Studio (M5 Max, 128GB)
For several departments that need smarter search and one login.
- For several departments
- A fast model and a reasoning model side by side
- Staff sign in with their existing work login
Technical: Qwen3.8-27B, Gemma 4 31B or Qwen3.5-122B-A10B, alongside Qwen3.6-35B-A3B, advanced document search, SSO, continuous batching
All-inclusive, no surprise costs. Includes:
- Machine setup, model installation, chat software and access control by department
- Loading all your internal documents within the assessed scope, including scanned documents that need OCR if identified in the assessment, and calibrating answer quality. The AI looks up your documents (RAG); the model is not retrained
- Sign-in with your existing company accounts (SSO)
- Testing, go-live and staff training
- 60-day warranty, then monthly maintenance and loading of new documents under the stated monthly fee
AI Enterprise
1 Mac Studio (M5 Ultra, 256GB, 512GB option)
Apple’s most powerful desktop chip, with enough memory for our largest open models on one machine.
- For company-wide use
- Our strongest model, can connect to your ERP
- Existing work login, data stays in your office
Technical: GLM-5.3-Flash (runs on its own, confirmed on your machine first) or Qwen3.5-122B-A10B alongside smaller models, 256GB unified memory (512GB option)
All-inclusive, no surprise costs. Includes:
- Machine setup, model installation, chat software and access control by department
- Loading all your internal documents within the assessed scope, including scanned documents that need OCR if identified in the assessment, and calibrating answer quality. The AI looks up your documents (RAG); the model is not retrained
- Sign-in with your existing company accounts (SSO)
- Testing, go-live and staff training
- 60-day warranty, then monthly maintenance and loading of new documents under the stated monthly fee
Connect it to the systems you already use
Link your private AI to your ERP, CRM, accounting or HR system, such as Odoo, SAP or software built in-house. Staff can then look up, summarise and draft from your live business data, with nothing leaving your office. Priced by what needs connecting.
A fixed price, with no surprise costs. The price agreed after the free assessment is fixed for the assessed scope (number and type of documents, number of users, systems to connect), and it does not increase after signing. We only charge more if you ask for new work outside the signed scope, such as another department, another system to connect, model fine-tuning or extra third-party software licences, and only with a written supplementary quote you approve before work starts.
AI models
The AI models we use, and what each is good at.
All are 2026 open models under the Apache 2.0 or MIT licence, free for commercial use, and run entirely on your own machines. We pick the model after the assessment.
| Model | Released | Type | Size (4-bit) | Intelligence index | Best at |
|---|---|---|---|---|---|
| Qwen3.6-35B-A3BQwen (Alibaba) · Apache 2.0 | Apr 2026 | MoE · 35B, 3B active | 20.4 GB | 18.2 | Fast answers for many people at once. Each extra user needs very little memory, so it suits busy teams. |
| Gemma 4 26B-A4BGoogle · Apache 2.0 | Mar 2026 | MoE · 26B, 4B active | 15.3 GB | 16.7* | Fast everyday questions and drafting, with a small memory footprint. |
| Gemma 4 31BGoogle · Apache 2.0 | Mar 2026 | Dense · 31B | 18.4 GB | 14.7 | Careful writing; suited to long documents. |
| Qwen3.8-27BQwen (Alibaba) · Apache 2.0 | Aug 2026 | Dense · 27B | 16.1 GB | 33.7 | Reasoning and following detailed instructions. Highest Artificial Analysis Intelligence Index score among the AI Box and AI Pro models. |
| Qwen3.5-122B-A10BQwen (Alibaba) · Apache 2.0 | Feb 2026 | MoE · 122B, 10B active | 69.6 GB | 17.7* | Broader knowledge for larger teams. Runs on AI Pro or AI Enterprise. |
| GLM-5.3-FlashZ.ai · MIT | Aug 2026 | MoE · 320B, 18B active | 204 GB | 41.8 | Our strongest model. Scores just above Gemini 3.8 Flash on the Artificial Analysis Intelligence Index (41.8 vs 40.9). Needs AI Enterprise; above our usual 75% memory rule, but published tests run it on the 256GB M5 Ultra, and we confirm it on your machine first. |
Intelligence index: Artificial Analysis Intelligence Index v4.3.2, best published setting (higher is better; * means estimated by Artificial Analysis). Size is the model file at 4-bit; every extra user needs a little more memory on top.
Other open models that run on the same machines
Beyond the models above, these open models are licensed for commercial use and fit on our machines. “Runs on” is the smallest package with enough memory: the 4-bit model file must fit within about 75% of the machine’s memory, leaving room for users. Larger packages run everything the smaller ones can.
| Model | Released | Type | Size (4-bit) | Runs on | Intelligence index |
|---|---|---|---|---|---|
| MiMo-V2.6-FlashXiaomi · MIT | Sep 2026 | MoE · 309B, 15B active | 167.3 GB | AI Enterprise | 37.9 |
| DeepSeek V4 Flash 0731DeepSeek · MIT | Jul 2026 | MoE · 284B, 13B active | 151.5 GB | AI Enterprise | 34.3 |
| Inkling-SmallThinking Machines Lab · Apache 2.0 | Jul 2026 | MoE · 266B, 12B active | 153.4 GB | AI Enterprise | 25.7 |
| Hy3Tencent · Apache 2.0 | Jul 2026 | MoE · 295B, 21B active | 170.3 GB | AI Enterprise | 25.3 |
| K2-Horizon-MoVA-36B-A4BInstitute of Foundation Models · Apache 2.0 | Sep 2026 | MoE · 36B, 4B active | 26.4 GB | AI Box | 25.3 |
| Ling-3.0-flash-VLinclusionAI · MIT | Sep 2026 | MoE · 124B, 5.5B active | 78.7 GB | AI Pro | 24.6 |
| Qwen3.5-27BQwen (Alibaba) · Apache 2.0 | Feb 2026 | Dense · 27B | 16.1 GB | AI Box | 22.9* |
| Qwen3.6-27BQwen (Alibaba) · Apache 2.0 | Apr 2026 | Dense · 27B | 16.1 GB | AI Box | 21.4 |
| Step 3.7 FlashStepFun · Apache 2.0 | May 2026 | MoE · 198B, 11B active | 110.8 GB | AI Enterprise | 19.5* |
| Muse Glimmer 30BMeta · Apache 2.0 | Aug 2026 | Dense · 30B | 19.4 GB | AI Box | 17.5 |
| GLM-4.7-FlashZ.ai · MIT | Jan 2026 | MoE · 30B, 3B active | 16.9 GB | AI Box | 14.9* |
| Granite 4.2 30BIBM · Apache 2.0 | Aug 2026 | Dense · 30B | 17.7 GB | AI Box | 14.8* |
| Gemma 4 12BGoogle · Apache 2.0 | Jun 2026 | Dense · 12B | 6.7 GB | AI Box | 14.2* |
| Command A+Cohere · Apache 2.0 | May 2026 | MoE · 218B, 25B active | 135.1 GB | AI Enterprise | 13.1 |
| Nemotron 3.5 LightningNVIDIA · OpenMDW-1.1 (permissive, MIT-style) | Aug 2026 | MoE · 30B, 3B active | 17.8 GB | AI Box | 12.9 |
| Seed-OSS-36BByteDance Seed · Apache 2.0 | Aug 2025 | Dense · 36B | 20.3 GB | AI Box | 12.1* |
| gpt-oss-120bOpenAI · Apache 2.0 | Aug 2025 | MoE · 117B, 5.1B active | 62.3 GB | AI Pro | 11.6 |
| Mistral Small 4Mistral AI · Apache 2.0 | Mar 2026 | MoE · 119B, 6.5B active | 67.8 GB | AI Pro | 11.3 |
| Granite 4.2 8BIBM · Apache 2.0 | Aug 2026 | Dense · 8B | 5.3 GB | AI Box | 11.1 |
| gpt-oss-20bOpenAI · Apache 2.0 | Aug 2025 | MoE · 21B, 3.6B active | 11.2 GB | AI Box | 9.0 |
| Devstral Small 2Mistral AI · Apache 2.0 | Dec 2025 | Dense · 24B | 15.1 GB | AI Box | 7.5 |
| Kimi Linear 48B-A3BMoonshot AI · MIT | Oct 2025 | MoE · 48B, 3B active | 27.6 GB | AI Box | 7.3* |
| Ministral 3 14BMistral AI · Apache 2.0 | Dec 2025 | Dense · 14B | 8.4 GB | AI Box | 6.0 |
| Phi-4-reasoning-vision-15BMicrosoft · MIT | Mar 2026 | Dense · 15B | 9.1 GB | AI Box | – |
| EuroLLM-22BUTTER project · Apache 2.0 | Dec 2025 | Dense · 22B | 12.7 GB | AI Box | – |
Sorted by Artificial Analysis Intelligence Index v4.3.2 (* estimated by Artificial Analysis; – not yet rated). Sizes are the 4-bit weight files on Hugging Face and do not include memory for users. Models close to the memory limit, or with new architectures, are tested on the real machine before we commit. Licences: Apache 2.0, MIT or OpenMDW-1.1, all allowing commercial use.
How they compare with commercial cloud AI
Scores on the Artificial Analysis Intelligence Index (higher is better). The best cloud models still lead. Our strongest model, on AI Enterprise, scores just above Gemini 3.8 Flash (41.8 vs 40.9), and your data stays in your office.
- Claude Opus 5.5Anthropic57.6
- GPT-6 AstraOpenAI52.7
- Gemini 4 ArgonGoogle52.6
- GLM-5.3-FlashOurs · AI Enterprise41.8
- Gemini 3.8 FlashGoogle40.9
- GPT-6 LunaOpenAI38.1
- Qwen3.8-27BOurs · AI Pro33.7
- Qwen3.6-35B-A3BOurs · AI Box18.2
- Gemma 4 26B-A4BOurs · AI Box16.7*
Source: Artificial Analysis Intelligence Index, accessed September 2026, best published setting for each model (* estimated by Artificial Analysis). Commercial models run in the provider’s cloud; ours run on your own Mac Studio.
What's included
Exactly what you get.
What we deliver
- Needs assessment, architecture design and rollout plan
- Hardware supply, installation and setup
- AI model installation and answer calibration
- A searchable knowledge base built from your documents (RAG), including OCR of scanned documents identified in the assessment
- Chat screen, admin panel, access by department and activity log
- Single sign-on with your existing login system (AI Pro and AI Enterprise)
- Testing, staff training and handover documents
- 60-day warranty and technical support after launch
What you provide
- Your documents and data, with the access we need
- Access to any systems you want connected
- Space, power and office network for the machines
- A business contact who works with us through the project
Only with a written supplementary quote
- New work outside the signed scope, such as another department
- Connecting another system not in the signed scope
- Fine-tuning a model for a specific writing style or domain
- Extra third-party software licences you ask for
How payment works
- Hardware
- Paid in full when ordered.
- Setup
- 50% at signing, 30% at user acceptance testing, 20% at launch.
- Maintenance
- Monthly, paid in advance, minimum 12 months.
Security commitments
Seven security commitments when we deploy private AI.
No system is perfectly secure. With AI, prompt injection cannot yet be fully prevented: the UK's National Cyber Security Centre writes that prompt injection attacks "may never be totally mitigated" the way SQL injection can be. So we reduce the risk in layers, and these are the seven we commit to on every deployment.
No outbound internet by default
The AI machine does not call outside AI services or APIs and opens no outbound internet port, except connections you approve, such as your login provider. Updates are loaded as digitally signed packages.
Permissions inside the document store
Each person can only look up the documents they are allowed to see. Permissions are filtered before the AI reads anything, so it never sees a document the user may not open.
The AI reads; people approve risky actions
AI tools run with the asking user’s own permissions, and risky actions need a person to approve them. An emergency stop switches the AI off at once.
Tamper-evident, append-only logs
A record of who asked what and which documents the AI read. The AI has no permission to edit or delete the log.
Hidden-instruction scan when loading documents
Documents are scanned for hidden text and injected instructions (prompt injection) before they enter the store. Answers never auto-load links or images from outside domains.
Sensitive data masked by role
ID numbers, salaries, bank account numbers and similar fields are shown only to authorised people; everyone else sees a masked version.
Departing staff data removed
A departing employee’s personal data is removed from the search store, caches and logs once there is no longer a legal or business reason to keep it, as section 25 of Singapore’s PDPA requires.
Singapore PDPA section 25: an organisation must cease to retain personal data, or anonymise it, once the purpose it was collected for is no longer served and retention is no longer necessary for legal or business purposes (Singapore Statutes Online).
Comparison
How it compares with ChatGPT, Copilot and Glean.
| Criteria | Public AI (ChatGPT, Copilot, Glean) | Namtech private AI |
|---|---|---|
| Where your data goes | The provider’s cloud, often overseas | Stays in your office |
| How you pay | Per person, every month | A fixed, predictable cost |
| Data protection | Depends on how it is set up | Built to keep data in-house |
| Knows your company documents | Limited | Yes, answers from your own files |
Implementation
Up and running in 8-10 weeks.
Five clear steps, each signed off before we move on. The exact time depends on how many documents you have and what we connect to.
- 01
1-2 weeks
Understand your needs
We agree what you need, how it will work and how we will measure success.
- 02
2 weeks
Set up the system
We install the machines, the AI models and the software.
- 03
2 weeks
Load documents and calibrate answers
We load your internal documents for the AI to look up (RAG) and calibrate answer quality. The model is not retrained.
- 04
1-2 weeks
Connect and test
We connect your logins, set who can see what, and test it with your team.
- 05
1 week
Launch and train
We go live, train your staff and hand over the documentation.
Industries
Who it is for.
Any team that works with private information can use it. A few examples, with all data kept inside your organisation:
Banking and finance
Find answers in policies, contracts and customer files, without data leaving the business.
Healthcare
Help staff find procedures and clinical guidance quickly, while patient records stay private.
Manufacturing
Let workers ask about procedures, manuals and maintenance guides on the factory floor.
Legal
Draft and review contracts and search past files, with confidential documents kept in-house.
Retail and e-commerce
Give staff one place to look up products, policies and orders.
Public sector
Search internal rules and procedures with data kept on your own premises.
Our commitments
What we promise.
Free assessment first
We start with a short, free review. You see a clear plan and price before you decide anything.
You own the equipment
The computers are bought through official channels and belong to your business, so you stay in control.
Warranty and ongoing care
A 60-day warranty after launch, then a monthly maintenance plan (minimum 12 months) to keep it running smoothly.
No lock-in
It uses open AI models (Qwen, Gemma, GLM) under the Apache 2.0 or MIT licence, so you are never tied to one vendor.
Pay as we deliver
Payment is split 50 / 30 / 20 across milestones, so you pay as each stage is approved.
Your data stays put
Every question and document is handled in your office, never sent to public AI services.
Glossary
Jargon, in plain English.
- Private AI platform
- An AI system that runs on equipment your business controls, so nothing you type leaves the company.
- On-premise
- Installed and run on machines in your own office, rather than on someone else’s cloud.
- RAG (Retrieval-Augmented Generation)
- The AI looks up the right pages in your documents before it answers, so its answers are based on your information and can show their source.
- MoE (Mixture of Experts)
- A type of AI model made of many small specialists. Each question only uses a few of them, so it gives good answers without needing huge computers.
- SSO (Single Sign-On)
- Staff use one work login for many apps, including the AI assistant.
- Data sovereignty
- Your data stays under your control and the rules of the place where it is stored.
Private AI FAQ
What is a private AI platform?
It is an AI assistant, similar to ChatGPT, that runs on computers in your own office. It answers questions using your company documents, and nothing you type is sent to outside AI companies.
Which AI models do you use?
Current open models released in 2026, under the Apache 2.0 or MIT licence: Qwen3.6-35B-A3B and Gemma 4 26B-A4B for fast everyday use, Gemma 4 31B for long documents, and Qwen3.8-27B for reasoning. AI Pro and AI Enterprise can also run Qwen3.5-122B-A10B, and AI Enterprise runs GLM-5.3-Flash, our strongest model. We choose the model after the assessment, based on your team size and documents.
What equipment does it run on?
Apple Mac Studio computers with M5 chips that sit in your office. There are three sizes: AI Box (M5 Max, 64GB), AI Pro (M5 Max, 128GB) and AI Enterprise (M5 Ultra, 256GB, with a 512GB option). Each is a single quiet desktop; the largest draws at most 385W according to Apple. No cloud, no outside AI service.
How many people can use it at the same time?
Many staff can share it, because only a few people are asking a question at any one moment. The serving software answers several questions together instead of one after another. In a published oMLX benchmark on a Mac Studio with M5 Max 128GB (jacar.es), Qwen3.6-35B-A3B answered 8 questions at once, with combined output of about 238 tokens (pieces of words) per second. We confirm the right size for your team during the assessment.
How do these models compare with ChatGPT, Claude or Gemini?
The best commercial cloud models still score higher. On the Artificial Analysis Intelligence Index, Claude Opus 5.5 scores 57.6 and GPT-6 Astra 52.7. Our strongest model, GLM-5.3-Flash on AI Enterprise, scores 41.8, just above Gemini 3.8 Flash at 40.9. The trade-off is that your data never leaves your office.
How is it different from ChatGPT or Microsoft Copilot?
ChatGPT and Copilot send your data to the provider’s cloud and charge for every person, every month. Our private AI keeps data in your office, costs a fixed amount to run, and answers from your own documents.
Will my data be used to train public AI?
No. Your data is only used inside your company, by the people who have access, and is never used to train public AI models.
Does it help with data protection rules?
Keeping data in your office lowers the risk of sending personal or confidential data overseas. That supports obligations such as the PDPA in Singapore and similar laws elsewhere. It is a technical safeguard, not legal advice.
How long does it take to set up?
Usually 8-10 weeks, in five steps: understand your needs, set up the system, load your documents, connect and test, then launch and train your team.
How much does it cost?
It depends on how many people will use it, how many documents it reads, the size of AI model and what it connects to. We give you a clear, fixed price after a free assessment.
What does the quote include? Are there extra costs?
Each package is all-inclusive for the assessed scope: machine setup, model installation, chat software and access control by department; loading all your internal documents in scope, including scanned documents that need OCR if identified in the assessment, and calibrating answer quality; sign-in with company accounts (SSO) on AI Pro and AI Enterprise; testing, go-live and staff training; and a 60-day warranty, then monthly maintenance and loading of new documents under the stated monthly fee. The agreed price does not increase after signing. We only charge more for new work outside the signed scope, such as another department, another system, model fine-tuning or extra third-party licences, and only with a written supplementary quote you approve before work starts.
Do you need to train the model?
No. Training a model from scratch needs enormous amounts of data and large GPU clusters; companies such as Alibaba (Qwen), Google (Gemma) and Z.ai (GLM) do this and release the trained models. We use these trained models and let the AI look up your company documents (RAG): for each question the system finds the right passages, and the model reads them to answer with sources. The model itself is not retrained, which is why the machine only needs enough power to run it. Fine-tuning a model for a specific writing style or domain is not part of the packages and is quoted separately on request.
Can it work with our existing systems and logins?
Yes. Staff can sign in with their existing work login (single sign-on), and it can connect to your ERP, CRM, accounting or HR systems. It also offers a standard OpenAI-compatible connection for your other apps.
How is private AI protected against attacks and data leaks?
No system is perfectly secure, and with AI, prompt injection cannot yet be fully prevented; the UK National Cyber Security Centre says such attacks may never be totally mitigated. We reduce the risk in layers, with seven commitments on every deployment: no outbound internet by default, with updates as signed packages; permissions enforced inside the document store before the AI reads anything; the AI reads only, with human approval for risky actions and an emergency stop; tamper-evident, append-only logs the AI cannot edit or delete; a scan for hidden instructions when documents are loaded, and no auto-loading of outside links or images; sensitive fields masked by role; and departing staff personal data removed from the search store, caches and logs once there is no legal or business reason to keep it, in line with section 25 of Singapore’s PDPA.
Is it licensed for commercial use?
Yes. It uses open AI models such as Qwen, Gemma and GLM under the Apache 2.0 or MIT licence, which allow commercial use, so you are not locked into one vendor.
A technical solution, not a substitute for legal advice.