AI Box
Your own AI, on hardware in your company
AI Edge Box and AI Office Box: local AI models on hardware in your company. All data stays in your network. Pick a package by what it has to run, from one camera on one station to large language models for the whole office.
Plug in and go
A box in your office, not a cloud subscription
The AI Box runs your models locally, on hardware in your company. It works without an internet connection, and we set it up, update it and look after it remotely over a secure connection.
- 1
Plug in the power
One normal socket. The box is about the size of a book and runs quietly on a desk or a shelf.
- 2
Connect it to your network
One cable to your router, or straight to your laptop. No server room and no cloud account needed.
- 3
Open it in your browser
Every laptop in your office uses the box like a website. Your data never leaves the building.
Packages
Seven packages, from one camera to large models
Edge Box works next to a production line; Office Box works in your office or server room. Approximate hardware prices, October 2026.
| Package | Hardware price (approx.) | For whom | Configuration |
|---|---|---|---|
| AI Edge Box Camera | PLN 3,000–5,000 per station | Quality control: 1 camera, 1 station | Jetson Orin Nano Super (8 GB) + industrial camera |
| AI Office Box Start | PLN 5,000 | 1 use case, a small model, up to 5 people | Ryzen 5 / Core i5, 32 GB RAM, 1 TB SSD, RTX 5060 Ti 16 GB |
| AI Office Box Pro | PLN 15,000–20,000 | Several models, vision, more users | Ryzen 7 / Core i7, 64 GB RAM, 2 TB SSD, RTX PRO 4000 (24 GB) |
| AI Edge Box Line | PLN 20,000–25,000 | Quality control: several cameras on a line | Jetson AGX Orin (64 GB) + industrial cameras |
| AI Office Box Pro+ | PLN 28,000–32,000 | All offers at once (language + vision + voice), up to about 20 people | Ryzen 7 / Core i7, 64 GB RAM, 2 TB SSD, 2× RTX PRO 4000 or RTX PRO 4500 (32 GB) |
| AI Office Box Max Lite | PLN 40,000–45,000 | 32B models, limited on-site coding | Ryzen 9, 128 GB RAM, 4 TB SSD, RTX PRO 5000 (48 GB), UPS |
| AI Office Box Max | from PLN 60,000 | Full on-site coding (KaizoCRM 24 in Full Local mode), large models | Ryzen 9 / Threadripper, 128 GB RAM, 4 TB SSD, RTX PRO 6000 (96 GB), UPS, encrypted NAS |
Edge Box and Office Box together
An AI Edge Box works next to each production line; an AI Office Box sits in your office or server room. At an industrial client, one AI Office Box Pro or Pro+ collects the results from every line, calculates KPIs and runs the language assistant.
The price is the hardware only
Each package is quoted as hardware. Implementation (configuration, model tuning, integration) and the maintenance subscription are added on top.
All data stays with you
Models run locally in your internal network. No data is sent to the cloud, and outgoing traffic is blocked on the firewall.
For each offer
Which AI Box runs which offer
KaizoCRM 24
AI Office Box Pro or Pro+ for Cloud Coding; AI Office Box Max (96 GB) for Full Local.
See the offer →KaizoReception
AI Office Box Start or Pro; the cheapest start is a 16 GB RTX PRO 2000 card.
See the offer →KaizoCraft AI Assistant
AI Office Box Start; the cheapest start is a 16 GB RTX PRO 2000 card.
See the offer →KaizoBooks AI OS for Accounting
AI Office Box Pro with a 24–32 GB GPU in your office; AI Office Box Start for a small number of statements. The demo runs on a laptop with a graphics card.
See the offer →KaizoMarketing AI
AI Office Box Pro with a 24–32 GB GPU for video, voice and image generation; AI Office Box Start if you only need texts and scheduling.
See the offer →KaizoCar 360
Capture: a phone or tablet with an Orbbec stereo depth camera (Gemini 2 or Gemini 335 class). Processing: AI Office Box Pro with a 24–32 GB GPU for the 3D models; AI Office Box Start for photos and the 360° spin only.
See the offer →KaizoFlow Supply Chain
AI Edge Box Camera at each dock or packing station, plus one AI Office Box Pro that collects the results, runs the language model and calculates the KPIs.
See the offer →KaizoVision Quality Control
AI Edge Box Camera or Line at each line, plus one AI Office Box Pro or Pro+ that collects results and calculates KPIs.
See the offer →KaizoCMMS
AI Office Box Start or Pro in the plant network; technicians use their phones (PWA), operators a tablet at the machine.
See the offer →
Graphics cards
How many models fit on each card
4-bit quantisation, approximate gross prices for the card only, October 2026.
| Card | Memory | Price (approx.) | Models | Example sets |
|---|---|---|---|---|
| RTX PRO 2000 Blackwell | 16 GB | PLN 4,800 | 2–4 small | KaizoReception: Bielik 11B + Whisper + TTS + embeddings. Quality control: YOLO + Qwen3-VL-8B |
| RTX 5060 Ti / 5070 Ti / 5080 | 16 GB | check current price | 2–4 small | As above (consumer cards: demos and proofs of concept) |
| RTX PRO 4000 Blackwell | 24 GB | PLN 10,000 | 3–4 | Mix: Bielik 11B + Qwen3-VL-8B + YOLO. Assistant: gpt-oss-20b + Whisper + embeddings. Or Gemma 3 27B alone |
| RTX PRO 4500 | 32 GB | PLN 17,700 | 4–5 | All offers at once: Bielik 11B + Qwen3-VL-8B + YOLO + Whisper + TTS. Or Qwen3-32B + YOLO + embeddings |
| RTX 5090 | 32 GB | PLN 20,000 | 4–5 | As RTX PRO 4500, faster (consumer card) |
| 2× RTX PRO 4000 | 2× 24 GB | PLN 20,000 | 4–6 | Card 1: gpt-oss-20b + Whisper. Card 2: Bielik 11B + Qwen3-VL-8B + YOLO |
| DGX Spark / GB10 | 128 GB shared | PLN 22,500 (whole computer) | 5+ | gpt-oss-120b + Qwen3-VL-32B + Bielik 11B + Whisper |
| RTX PRO 5000 | 48 GB | PLN 32,600 | 4–6 | Qwen3-32B + Qwen3-VL-8B + Whisper + YOLO. Or Qwen3-VL-32B + Gemma 3 27B |
| RTX PRO 6000 Blackwell | 96 GB | PLN 48,600 | 5+ | gpt-oss-120b + Qwen3-VL-8B + Whisper + YOLO. Or Qwen3-32B + Qwen3-VL-32B + Bielik 11B with room for many users |
Our recommendation: Best value for an MVP: RTX PRO 4000 (24 GB) at about PLN 10,000, with low power use and room for 3–4 models. The cheapest start for KaizoReception or basic quality control: RTX PRO 2000 (16 GB) at about PLN 4,800. All offers on one machine: RTX PRO 4500 (32 GB) at about PLN 17,700.
Memory per model
| Model | Used for | Memory (4-bit) |
|---|---|---|
| YOLO | Defect detection | about 1 GB |
| Whisper turbo | Speech to text | about 2 GB |
| TTS (e.g. Piper) | Text to speech | about 1 GB |
| Embeddings | Search in documents | about 1 GB |
| Qwen3-VL-8B | Vision + description | about 6 GB |
| Bielik 11B | Polish language model | about 7 GB |
| Qwen3-14B | Language model | about 9 GB |
| gpt-oss-20b | Language model | about 13 GB |
| Gemma 3 27B | Language model | about 17 GB |
| Qwen3-32B / Qwen3-VL-32B | Language / vision | about 20 GB |
| gpt-oss-120b | Large language model | about 65 GB |
Good to know
What we check before every quote
A card is not a computer
GPU prices are for the card only. Add a workstation (CPU, RAM, disk, power supply) at about PLN 8,000–15,000. DGX Spark is a complete computer.
Users take memory too
Every parallel conversation needs room for its context. On 16 GB, a few simultaneous users; on 48 GB and more, several dozen. Keep about 10–15% of memory free.
DGX Spark
Fits the most models, but is slower than RTX cards on large models (lower memory bandwidth). ARM architecture: check support for libraries and camera SDKs.
Real-time vision
Many cameras and a fast line need compute, not just memory. With several cameras, use a separate card for vision only.
Professional vs consumer cards
For 24/7 work at a client: PRO cards (ECC memory, lower power, quieter). Consumer cards for demos and proofs of concept.
Models and prices change
Model names are examples of open model families; before a rollout we check newer versions and commercial licences. GPU prices changed monthly in 2026, so every quote is checked with the distributor.
Not sure which AI Box you need?
Tell us what it has to run and for how many people. We pick the package and check today's prices with the distributor.
Request MVP