IT firms, lawyers, notaries
An AI server in your office. What you give it to read never leaves the building.
You get the hardware, an open-weight model and the support. The box is sealed and has no path to the internet. One monthly fee, fully managed.
Request a 20-minute callWe test the box in our workshop before delivery, with you or your IT person watching. The prepay can sit in an escrow account, not with us.
One box serves up to
- 60people through the day
- 8ktokens of context
- 30/stokens a second each (p90)
The Office box, fp8 KV cache, thinking off. With thinking on: up to 40 people, on the same terms. Measured on a rented card limited to the memory of the Office card. It is an estimate until we measure our first box.
The fee does not move with tokens, agents or API calls.
What it costs against the cloudWhere the data goes
Which document or code would you never send to a cloud model?
No law we have read bans cloud AI. At ordinary volume, cloud AI is cheaper and the right choice. Bastion is for what you would never send.
What GDPR and DORA ask of you01Inside your walls.
Your people ask from their desks; the box answers inside the building.
MeasuredCCBE, Technical guide on the use of AI tools and models by lawyers, page 8 (opens in a new tab)Read 2026-09-26
02Outside, 0 bytes.
The box cannot reach the internet. The one path out stops at the wall.
MeasuredMicrosoft Learn, Data, privacy and security for Foundry Models sold by Azure (opens in a new tab)Read 2026-09-22
03The gate stays open.
You or your IT person watch the acceptance test in our workshop, before delivery.
Contract termHow do payment and the acceptance test work?
04A bastion you can see inside.
Each step above shows its source. The Evidence section shows the date we read each document.
What you get
One box, not a project.
Everything below is in the fee.
- The box
- A professional GPU computer, sized to your team's work.
- The model
- Qwen3.8-27B, an open-weight model under the Apache 2.0 licence. It drafts, summarises and reviews documents and code, in Romanian and English. It reads text: Word, email, editable PDF. It does not read scanned pages.
- Only on your network
- The box answers on your network only. Updates arrive as signed bundles.
- A swap, not a repair
- If the box fails, we replace the whole box.
How we work
Four steps, each one small.
20 minutes
You tell us which data should never leave. We size the box.
Prepay or escrow
At signing you pay 6 months of the fee, direct or into an escrow account. The standard term is 36 months.
The workshop test
The box passes its acceptance test before it leaves our workshop. The pass marks are in the questions below.
A socket and a network cable
You plug in the box. Your team works in a browser, and your apps connect through the API.
Evidence, not adjectives
Check us before you trust us.
Every claim here carries a label, a source and the day we read it.
- Measured
For a Global deployment, prompts and responses may be processed in any geography where the model is deployed.
Microsoft LearnData, privacy and security for Foundry Models sold by Azure (opens in a new tab)Read 2026-09-22
- Measured
A provider may also promise deletion ... but they can still be compelled by mandatory legal provisions to retain or hand over data.
CCBETechnical guide on the use of AI tools and models by lawyers, page 8 (opens in a new tab)Read 2026-09-26
IT and software firms
The agents run on your developers' computers, and the model runs on your box. The model is Qwen3.8-27B, not Claude Opus.
Why the vendor's perimeter is not yoursLaw firms, notaries and small firms
Research, contract review and drafting on client files. We size the box by the length of your documents.
Private AI for law firmsBanks and insurers
What DORA and the national bank ask of an AI supplier.
DORA and AIHealth and public bodies
What you can run on patient data, and on the data a public body holds.
Private AI in healthcare
Frequently asked questions
On-premise AI, answered plainly.
How much does it cost compared to ChatGPT or Claude?
At ordinary volume, cloud AI is cheaper. Bastion is one monthly fee that does not move with use. Against a frontier cloud model, a full Reasoning box breaks even at about 130,000 tokens a person a working day. Below that, the cloud is hard to beat. Your offer states one fee for your setup.
How many people can use one box?
It depends on the box and on the thinking mode. The Office box serves up to 60 people with thinking off, or 40 with thinking on. The Reasoning box serves up to 110 people with thinking off, or 70 with thinking on. Both at 8k tokens of context, 30 tokens a second each (p90), with an fp8 KV cache. For coding agents, one card serves up to 4 agents at 64k tokens of context, at the same speed and cache, thinking off. People means named people who use the box through the day. We measured these figures on a rented card. On our box they are an estimate until we measure our first box.
Which model runs on the box, and is it as good as GPT or Claude?
Qwen3.8-27B, a dense open-weight model under the Apache 2.0 licence. It handles drafting, summaries, document review, coding help and chat in Romanian or English. Frontier cloud models still lead on the hardest reasoning and coding tasks. Many teams use both: the cloud for work that may leave, Bastion for work that must stay.
How do payment and the acceptance test work?
You pay 6 months of the fee at signing, direct or into an escrow account. They pay for months 1 to 3 from delivery and months 34 to 36 of the 36-month term. Before delivery, the box passes an acceptance test in our workshop, on our test set, with you or your IT person watching. No document of yours enters the test. Our first three Reasoning boxes pass at 90 people with thinking off and 70 with thinking on. Both at 8k tokens of context, 30 tokens a second each (p90), with an fp8 KV cache. If a box fails, we get one 15-day fix period. The contract says what happens to the prepay if the fix fails.
Does it work without internet?
Yes. The box answers on your local network with no connection to the internet. An optional battery carries a short power cut, then shuts the box down safely. It is not a generator.
Is Bastion GDPR compliant?
No box makes an organisation compliant by itself. Compliance depends on your processing, your legal basis, your policies and your access controls. Bastion keeps the AI processing inside your premises, so the data does not go to a cloud provider. We supply the box and the documents that support your work.
Does Bastion support DORA?
Each box will ship with an audit file: the contract annex written to DORA Article 30, the register row, the exit plan and the acceptance test record. DORA compliance stays with the regulated entity and depends on the use case and the contract.
A short call
Request a 20-minute call.
After the call, you get in writing, within 5 working days, which box fits you, or why none does.
Prefer to size the box yourself? Open the configurator- Or write to contact@bastion-inc.com
- Bastion, Cluj-Napoca, Romania