AI with no internet connection. Nothing leaves your building.
The offline edition runs the language model on your own GPU server. The software goes online once during setup to download what it needs. After that it works with no internet at all.
Complete data privacy
Prompts, documents and answers never leave your network. Ideal for defence, finance, healthcare and the public sector.
Open models on your GPU
A built-in llama.cpp engine runs gpt-oss, Qwen, Gemma and other open models. All of them are free for commercial use.
Checked before you go offline
Before the connection is cut, a test confirms that the model uses its tools correctly and a self-check confirms that everything works.
Current models, local and cloud
The catalog is updated with new releases. Local models run on your own hardware; cloud models need an internet connection and a provider key.
| Family | From | Sizes |
|---|---|---|
| gpt-oss | OpenAI | 20B, 120B |
| Qwen3.8 / Qwen3.6 / Qwen3.5 | Alibaba Qwen | 4B, 9B, 27B, 35B A3B |
| Qwen3-Coder-Next | Alibaba Qwen | 80B A3B |
| Gemma 4 | E4B, 12B, 26B A4B | |
| Granite 4.2 | IBM | 3B, 8B, 30B |
| Ministral 3 | Mistral AI | 8B, 14B |
| Phi-4 | Microsoft | mini, 14B |
| Qwen2.5-VL, MiniCPM-V | Alibaba Qwen, OpenBMB | 3B, 7B, 8B (vision) |
| BGE-M3, Nomic Embed | BAAI, Nomic | document search |
| Provider | Models |
|---|---|
| Anthropic | Claude Haiku 4.5, Sonnet 5.5, Opus 5.5, Fable 5.1 |
| OpenAI | GPT-4.1, GPT-4.1 mini, GPT-4o mini |
| Google free tier available | Gemini 3.5 Flash, Gemini 2.5 Pro, Flash, Flash-Lite |
| Groq free tier available | gpt-oss 20B, 120B |
| Cerebras free tier available | gpt-oss 120B, GLM 4.7 |
| Mistral AI free tier available | Mistral Small, Ministral 8B |
Already running a model server? DepartmentAssistant can also connect to vLLM, Ollama or LM Studio on your network.
Pick the tier that fits your company
More GPU memory means a smarter model and longer context.
| Tier | GPU | Model | RAM | Free disk |
|---|---|---|---|---|
| Small | 1 × 24 GB (RTX 4090, A10, L4) | Qwen3.8 27B | 32 GB | 60 GB |
| Small+ | 1 × 32 GB (RTX 5090) | Qwen3.6 35B A3B | 48 GB | 60 GB |
| Medium Recommended | 1 × 80–96 GB (H100, RTX PRO 6000) | gpt-oss 120B, Qwen3-Coder-Next | 128 GB | 150 GB |
| Large | 2–8 × data-centre GPUs | GLM-5.3, DeepSeek V4 (vLLM) | 512 GB | 1 TB |
Large tier: very large models need their own GPU server, which runs Linux and the free vLLM software. DepartmentAssistant stays on your Windows server and connects to that model server over your internal network. We set this up together with you.
Online once, offline from then on
Install the engine
The app detects the GPU and installs the matching engine build.
Download the model
Choose the model for your hardware tier and activate it.
Prepare for offline use
One click downloads the engine, models, Python packages and OCR data.
Switch offline mode on
The self-check confirms that everything works. Outgoing connections are blocked from then on.
Exceptions you control
Need one outside service? It can be opened as a named exception, set up with our support. Every connection is logged.
- Microsoft 365: Outlook mail and Teams notifications
- Safe web search, results only
- Every other connection stays blocked
Try the offline edition
The demo includes the offline edition. Test it on your own GPU machine or a rented one.