Independent company · we are not affiliated with, sponsored by, or endorsed by the Swiss AI Initiative (EPFL, ETH Zurich, CSCS), the developers of the Apertus model · Official site: apertus-ai.org
Local AI hardware
AI Box and AI Powerhouse bring open-weight models, your apertus.ai gateway and Open WebUI into your building: delivered preconfigured, maintained by us. Your data and databases stay inside your network.
AI Box · NVIDIA GB10 · 128 GB
01 Two sizes
Both systems ship with the gateway, Open WebUI and selected models preinstalled. You plug them in, we take care of updates, monitoring and maintenance.
For teams
Nº 01
Compact, quiet and ready to run.
from€5,900
depending on configuration, excl. VAT
For departments & companies
Nº 02
More users, larger models, on-premise.
On request
depending on GPU count and build
02 Benchmarks
Indicative vLLM figures for current open-weight models: the speed of a single request, how many requests run smoothly in parallel and how many employees a system realistically serves day to day.
Reading speed ≈ 6 tokens/s
AI Box · NVIDIA GB10 · 128 GB
Gemma 4 26B-A4B FP8 + MTP
Smooth in parallel8 requests at 80 tokens/s
Servesup to ~25 employees
Qwen3.8 27B NVFP4 + MTP
Smooth in parallel10 requests at 10 tokens/s
Servesup to ~10 employees
Powerhouse 1× · 1× RTX PRO 6000 · 96 GB
Qwen3.8 27B FP8 + MTP
Smooth in parallel32 requests at 30 tokens/s
Servesup to ~60 employees
Gemma 4 31B NVFP4
Smooth in parallel4 requests at 23 tokens/s
Servesup to ~30 employees
Qwen3.8 27B FP8 + MTP · coding agents, 32k context
Smooth in parallel6 requests at 30 tokens/s
Serves~6 devs with coding agents
Powerhouse 2× · 2× RTX PRO 6000 · 192 GB
Qwen3.8 Flash-Next FP8
Smooth in parallel32 requests at 23 tokens/s
Servesup to ~100 employees
Qwen3.8 27B FP8 + MTP · 2 replicas
Smooth in parallel64 requests at 30 tokens/s
Servesup to ~120 employees
Powerhouse 4× · 4× RTX PRO 6000 · 384 GB
GLM-5.3-Flash NVFP4 + MTP
Smooth in parallel8 requests at 110 tokens/s
Servesup to ~150 employees
Qwen3.8 Flash-Next FP8 + MTP
Smooth in parallel32 requests at 20 tokens/s
Servesup to ~120 employees
Qwen3.8 27B FP8 + MTP · 4 replicas
Smooth in parallel128 requests at 30 tokens/s
Servesup to ~200 employees
Per system, the model from the table that serves the most employees · one dot = 20 employees
Indicative figures as of October 2026 from public vLLM benchmarks by independent labs and the community on the same hardware, rounded down. MTP means the model writes several tokens per step with its own prediction head, without losing quality. Smooth in parallel: measured concurrent requests with the speed each of them gets, mostly with short prompts; long documents lower the figures. Employees: a cautious estimate with 10–15% of the staff active at the same time and a quarter of the capacity kept free for agents and long documents. Before delivery we measure with your workloads.
03 On-premise
When apertus.ai runs on your premises, the MCP gateway connects local models directly to your internal databases, ERP, DMS and file shares. Every access goes through permissions and the audit log, and no request ever leaves your network.
Your network
Your data sources
apertus.ai gateway
AI & MCP, on-premise
04 Always included
Gateway, Open WebUI and your models are set up when the system arrives.
We choose models for your use cases, user numbers and languages and keep them up to date.
We handle monitoring, security and model updates, on site for the AI Powerhouse.
Through the gateway you combine local models with EU cloud models, for example for peak load.
The MCP gateway connects the models securely to your databases, ERP and DMS.
Hardware, software and models from one provider, with a named contact on our side.
FAQ Questions about local AI
Open-weight models up to around 200B parameters. The fast MoE model Gemma 4 26B-A4B suits several users at once; Qwen3.8 27B handles demanding tasks and coding. We choose the models that fit your use cases together with you.
The AI Box is enough for teams of up to around 25 people in everyday chat. For whole departments, coding agents or very large models, the AI Powerhouse is the right choice. We work it out with your numbers.
Yes. In the gateway you decide which requests stay local and which may use cloud models in the EU, for example for peak load or very large models.
We monitor the systems and roll out security and model updates through a secured access you control. For the AI Powerhouse we also come on site for installation and maintenance.
The AI Box is compact and runs in the office or a server cabinet. The AI Powerhouse belongs in a server room or your data centre, as a tower or 19-inch server.
Contact 20 minutes, an honest assessment
Book a short call. We look at your situation and tell you honestly whether and how apertus.ai can help.
What happens in the call