An AI server that works without an internet connection.
Tokeno is delivered as a ready server: hardware, software, a language model and documentation. It is installed on your premises, and your staff use it in a browser over the internal network.
What is in the box
- Hardware
- A server sized for your number of users and your use. The size classes are in the table below.
- Software
- An open source software stack apart from NVIDIA’s driver and CUDA runtime, with a hardened configuration.
- Language model
- The best model that fits the size class you bought. We keep it current.
- Admin view
- Users and access rights, the audit log, the state of network connections and software licences in one view.
- Documentation
- Administrator’s guide, installation instructions and the bill of materials (SBOM).
How it works
- InstallationThe server is installed in your server room or equipment room and connected to your internal network.
- UseStaff use the server in a browser over the internal network. Accounts and access rights are managed in the server’s own admin view.
- DataQuestions, documents and answers are processed on the server. They do not leave the building.
- UpdatesUpdates arrive as signed packages on separate media, not over the internet.
How the check works
The open software base has connections to outside services by default: web search, telemetry, external text recognition, a cloud code interpreter, a plugin store and external tools. They are closed when the server is built, not left to the customer’s settings.
- The check does not trust settings. The server tries to open every connection that the closed features would have used.
- Each attempt is recorded with one of three results: the connection opened, the connection did not open, or the result could not be confirmed.
- An unconfirmed result is not counted as closed. A path counts as closed only when the connection was attempted and did not open.
- The result is shown in the server’s admin view. You can run the check at any time.
You can confirm the same with your own tools: your firewall log shows whether the server tries to connect to the outside.
Size classes
There are four size classes. The class is set by how much GPU memory is available.
| Class | GPUs | Usable GPU memory |
|---|---|---|
| Small conditional / availability to be confirmed | 1–2 | about 32–64 GB |
| Workstation | 2–3 | about 190–290 GB |
| Rack | set in the quotation | about 600–900 GB |
| Data centre | set in the quotation | 1.4 TB or more |
Configuration, model and availability are always confirmed in the quotation.
You get the best language model that fits the size class you bought, and under the service agreement we keep it current.
What the delivery includes
- Bill of materials
- Every version ships with a machine-readable bill of materials (SBOM, CycloneDX format). It lists every software component with its origin and licence.
- Licence texts
- Licence and copyright information for every open source component. The admin view has its own page for them.
- The whole stack
- The software, configuration and documentation are yours.
- Exit clause
- The contract has an exit clause: you can leave without licence fees and keep the whole stack, configuration and documentation.
- No subscriptions
- Nothing requires a subscription to a third party.
Service contents
Six things you do not have to do yourselves.
| Service | What it means |
|---|---|
| Regular updates | Updates to the software base every quarter, tested before installation. |
| Updates without an internet connection | Signed update packages in which the model, runtime and driver have been tested together for your server. |
| Audit reports | The integrity of the audit log is checked on a schedule, and you receive a signed report of the result. |
| Finnish quality evaluation | Answer quality is measured against a Finnish-language evaluation set after every model or version change. |
| Model licence tracking | Model terms of use can change without notice. We track them so that no model whose terms conflict with your contract ends up on the server. |
| Support and response time | The response time is written into the contract. |
Get in touch.
We’ll tell you which configuration fits your use and what the delivery includes.
Phone: +358 44 973 8676