↳ Local AI

Private AI, on your own hardware.

Cloud AI transmits your data to a third party's servers. Local AI runs on infrastructure you own, on your premises. Nothing leaves the building. We select the model, specify the hardware, handle the installation, and train your team.

↳ Bound by secrecy

If you can't paste it
into ChatGPT.

Wealth managers, fiduciaries, lawyers, doctors, anyone bound by professional secrecy (Art. 321 StGB), cannot send client data to a cloud model hosted abroad. A model that runs on your own premises gives you the same AI leverage while the confidential data never leaves the building.

Built for regulated Swiss work: Vermögensverwalter · Fiduciaries · Legal · Healthcare · Finance
001 Local vs cloud

Same answers.
Different trust.

Cloud AI is operated by a third party on infrastructure you don't control. Local AI runs on infrastructure you own. For organisations handling sensitive, regulated, or confidential data, where the model runs is the deciding factor.

Dimension
Local · on-prem
Cloud · API
Where data lives
→In your building. Always.Never transmitted
On a vendor's servers, often abroad.Cross-border transfer
Cost model
→One-time hardware + setup.No per-token bill
Pay per token, every request, forever.Scales with usage
Compliance
→Air-gappable. Audit-friendly.Health · legal · finance
Depends on the provider's terms & region.Shared responsibility
Connectivity
→Operates fully offline.No internet required
Requires a live connection per request.Dependent on provider uptime
Control
→You own the model & the box.No silent changes
Model can change or retire without notice.Vendor's roadmap
Best suited to
→Privacy and predictable cost.Steady, sensitive workloads
Access to the absolute frontier model.Variable, non-sensitive tasks
Our recommendation: Frontier cloud models remain the most capable overall. · For confidential data, a capable model you own outperforms a stronger one you rent.
002 Models for on-premisereviewed November 2026

One default,
a full range.

For most clients our go-to is Qwen 3.8: on the independent leaderboard it sits in the same tier as models ten times its size, and it runs on a single workstation. Need lighter? We scale down to Gemma 4. Need the absolute ceiling? We scale up to DeepSeek or Kimi. All open-weight, all yours, with 4-bit memory figures shown.

Our go-to
Qwen 3.8 Alibaba
Apache 2.0 · unrestricted commercial use

A 27-billion-parameter model built for reasoning, coding and long multi-step tasks, that reads text, images and video. It runs on one workstation instead of a server. The default we deploy unless your needs push smaller or bigger.

27B · 4-bit workstation · 17 GB
27B · 16-bit workstation · 56 GB
↓ Scale down · smaller
Gemma 4 Google
Apache 2.0 · unrestricted commercial use

When a workstation is overkill. Google's efficient, multimodal open family runs on a laptop: capable enough for most everyday tasks on hardware you likely already have.

  • E4B Smallest footprint, 4-bit laptop · 5 GB
  • 12B Lightweight assistant, 4-bit laptop · 7 GB
↑ Scale up · bigger
DeepSeek · Kimi Heavyweights
MIT · Kimi K3 licence · commercial use OK

When you want the absolute ceiling. Models of more than 500 billion parameters for the highest-stakes reasoning, both in the top tier of the independent leaderboard. These need dedicated server hardware.

  • DeepSeek V4.1 Flash 552B, top tier, 4-bit server · 276 GB
  • Kimi K3 2.8T, top tier, 4-bit cluster · 1,400 GB
Sizing & rankings: tiers and server memory per the Onyx Self-Hosted LLM Leaderboard · laptop and workstation memory is the size of the published model files · final model selection is part of the audit.
003 Private install · pricing

Installed,
not rented.

Fixed-price engagements to put a private model into production on your premises. Hardware is procured by you to our specification, so it is yours from day one. Choose the depth of engagement you need.

Tier 01

Setup

Up and running

CHF 3'500
  • We spec the hardware, you procure it
  • Model installed & tuned on your box
  • Private chat interface for your team
  • Runs fully on-premise, offline-ready
  • Workflow audit
  • Team training
  • First agent built
Tier 03

Automate

Adopt + first agent live

CHF 7'500
  • Everything in Adopt, plus:
  • One agent built for your workflow, live in production
Every engagement: Runs on-premise · Open-weight models you own · Hardware to our specification, procured by you · No per-token fees
[ Next step ]

Get in touch now.

Get in touch→