On-premise AI box
The generative-AI box that stays in your walls
Compact, sealed, pre-loaded. Apple Silicon or Linux x86 depending on your datacenter. 100% open-source stack. No outbound connection by default.
Silicon
Apple M3 Ultra
Unified RAM
64–192 GB
Network
10 GbE
Air-gap
Compatible
Three sizes
Pick your fit
gbox S
s
For teams of 20-40 people
RAM
64 Go
Concurrent users
up to 40
Throughput
≈ 28 tok/s
Power
65 W
12 000 €
See details →
★ Most popular
gbox M
m
For mid-market organisations of 50-150
RAM
96 Go
Concurrent users
up to 150
Throughput
≈ 52 tok/s
Power
180 W
25 000 €
See details →
gbox L
l
For 150+ users or multi-site
RAM
192 Go
Concurrent users
up to 400
Throughput
≈ 85 tok/s
Power
350 W
38 000 €
See details →
Inside the box
A 100% open-source stack, in 6 layers
No proprietary component. You can audit, harden and extend. Everything is documented and installable on your box.
06
User interface
A familiar ChatGPT-style UI, accessible with corporate credentials. Multi-language, multi-model, multi-module.
Open WebUI
05
Model gateway
Smart routing to the right model per task. Quotas, metrics and logging centralised.
LiteLLM
04
Inference runtime
Runs open-source models locally on Apple Silicon or NVIDIA GPUs. Automatic quantisation per box size.
Ollama · llama.cpp
03
Document index
Postgres + vector extension. Embeddings computed locally on connected sources (SharePoint, Drive, etc.).
Postgres + pgvector
02
Enterprise authentication
SSO with Active Directory, Azure AD, Okta, Google Workspace via OIDC/SAML. Fine-grained RBAC, detailed audit logs.
Authentik
01
Hardened host OS
macOS on Apple Silicon or Ubuntu LTS on x86. Signed updates, minimal services, restrictive firewall by default.
macOS / Ubuntu LTS
Data flow
How a request travels (and where it doesn't)
The path of a question, from the user's keyboard to the answer — entirely inside your LAN.
User
Enterprise SSO
gbox
Doc index
Sourced answer
No outbound connection by default
No third-party API, no telemetry, no cloud model called in the background. The box can run 100% disconnected (air-gap mode).
What you get
The box, and everything that goes with it
The sealed box
Compact, pre-loaded with models, ready to plug in. Serial number + SHA256 signature provided.
Cables and PSU
10 GbE network cable, power supply, Thunderbolt cable for initial config.
Install runbook
Step-by-step printed + digital guide. Procedures for Azure AD, Okta, Google Workspace SSO.
Onboarding session
3 days on-site with a gbox engineer. User training included for your champions.
Admin dashboard included
One screen to drive your box
Health of all 18 services, connector sync timelines, searchable audit logs, Restic backups, multi-tenant orgs, GDPR right-to-erasure — all in a single interface, shipped by default.
https://gbox.local/admin/
Services up
18 / 18
Connectors
14
Last backup
2 h
Disk free
68 %
ollama
ok
litellm
ok
open-webui
ok
authentik
ok
prometheus
ok
grafana
ok
loki + vector
ok
gbox-orchestrator
ok
Real-time health
State of every service with timestamps, HTTP codes, latency metrics. Auto-detect drift.
Connector sync
Per-connector timeline, last sync, doc count. One-click manual sync trigger.
Searchable audit
Full-text search over every event (Loki + Vector). SIEM-ready export (Splunk, Elastic, Wazuh).
GDPR right-to-erasure
Erase a user via a confirmation-phrase form. Exhaustive trace of every action taken.
Ready to see the box in your context?
30-minute demo on your priority use cases. Paid 30-day on-site PoC, fully credited towards the contract.