Unit 01
onlineml1
Draw250 W / 400 W
- GPU
- NVIDIA A100 PCIe
- VRAM
- 40 GB HBM2
- CPU
- 2× AMD EPYC 7343 · 32 cores
- System memory
- 257 GB
The bench
Where we test, before we recommend anything.
These two A100 servers in Amsterdam are our proving ground — not the limit of what we can run for you. We put a candidate model on them against your own data first, so a recommendation is something we have watched work, not a number from someone else's slide. Whatever your project actually needs is specced and provisioned to fit it: rented in the EU by the hour, or bought outright, sized to the job.
Unit 02
onlineml2
Draw300 W / 400 W
- GPU
- NVIDIA A100 PCIe
- VRAM
- 80 GB HBM2e
- CPU
- 2× AMD EPYC 74F3 · 96 threads
- System memory
- 257 GB
The bench is not the ceiling. When a project needs a bigger card, or many, we size it and provision it — inference still runs on hardware we control, and nothing is forwarded to a third-party model provider.
The market we choose from
Open models, by weight and by what you may serve.
Which model suits a given job is decided on the use-case pages, from independent benchmarks. This is the shortlist those pages draw from: how much memory each model’s weights take, and — the part a licence hides behind its name — whether a company in the EU may serve it at all.
Weight at 4-bitMay not be served in the EUDashed line: a card class (GB)
Table view
| Model | 4-bit | Context | Licence | May we serve it |
|---|---|---|---|---|
| Qwen3-4B-Instruct-2507 | 2 GB | 256K | Apache-2.0 | Yes, freely |
| Teuken-7B-instruct-v0.6 | 3.9 GB | 4K | CC-BY-NC-4.0 | No |
| Qwen3-8B | 5 GB | 32K | Apache-2.0 | Yes, freely |
| Gemma 3 12B | 6.7 GB | 128K | Gemma Terms of Use | Yes, with conditions |
| Phi-4 | 7.8 GB | 16K | MIT | Yes, freely |
| Qwen3-14B | 8 GB | 32K | Apache-2.0 | Yes, freely |
| Mistral Small 3.2 (24B Instruct) | 13.4 GB | 128K | Apache 2.0 | Yes, freely |
| Gemma 3 27B | 15.1 GB | 128K | Gemma Terms of Use | Yes, with conditions |
| Qwen3-30B-A3B-Instruct-2507 | 17 GB | 256K | Apache-2.0 | Yes, freely |
| Granite 4.0 H-Small | 17.9 GB | 128K | Apache-2.0 | Yes, freely |
| Qwen3-32B | 18 GB | 32K | Apache-2.0 | Yes, freely |
| Llama 3.3 70B Instruct | 39.2 GB | 128K | Llama 3.3 Community License Agreement | Yes, with conditions |
| Hunyuan-A13B-Instruct | 44.8 GB | 256K | Tencent Hunyuan Community License Agreement | No |
| GLM-4.5-Air | 59.4 GB | 128K | MIT License | Yes, freely |
| Command A | 62.2 GB | 256K | CC-BY-NC-4.0 | No |
| Mistral Large 2 (2411) | 68.9 GB | 128K | Mistral AI Research License (MRL) | No |
| Qwen3-235B-A22B-Instruct-2507 | 132 GB | 256K | Apache-2.0 | Yes, freely |
| GLM-4.6 | 199.9 GB | 200K | MIT License | Yes, freely |
| Llama 4 Maverick 17B-128E Instruct | 224 GB | 1024K | Llama 4 Community License Agreement | No |
| DeepSeek-V3.2-Exp | 375.8 GB | 128K | MIT License | Yes, freely |
| Mistral Large 3 (675B Instruct) | 378 GB | 256K | Apache 2.0 | Yes, freely |
| Kimi-K2-Instruct | 560 GB | 128K | Modified MIT License | Yes, with conditions |
| DeepSeek-V4-Pro | 896 GB | 1024K | MIT License | Yes, freely |
The building
NorthC, Amsterdam.
Our servers are not under a desk. They sit in NorthC’s Amsterdam datacentre, a Dutch operator running Tier 3 facilities across the country. That gives the machines redundant power, cooling and connectivity, and gives you a physical address in the EU for the hardware your data runs on.
- Operator
- NorthC Datacenters
- Site
- Amsterdam
- Country
- Netherlands
- Facility standard
- Tier 3
Whose certificates these are.
NorthC holds ISO 27001, ISO 9001, ISO 14001, ISO 22301 and PCI DSS for its facilities. Those belong to the datacentre, not to ML Env. A certified building tells you the power, access control and continuity are audited; it says nothing about how any tenant writes software. We mention it because it is relevant, and we draw the line because it is honest.
NorthC certifications ↗Where your data goes
It stays in the Netherlands.
The honest reason to self-host is a narrow one, and it is worth more than it sounds. Your customers' data stays on servers in the EU instead of being sent to an American AI company. That removes the single most fragile step in the chain. It does not make you compliant on its own — you still need a lawful basis, the right notices, and decent security — but it takes one genuinely unstable problem off the table.
- No transfer to a third country
- Sending customer data to a US model provider is an international transfer under GDPR. It needs an adequacy decision or standard contractual clauses and an assessment behind them. Keeping inference here removes that step entirely.
- Nobody else keeps a copy
- There is no external processor holding your inputs. Several US APIs retain them for around thirty days, and some train on them unless you opt out.
- Not a compliance certificate
- Hosting location is one leg of the analysis. You still owe your customers a lawful basis, a notice that they are talking to a machine, and a data-protection assessment where the processing warrants one. We will not pretend otherwise.
- The AI Act does not change with hosting
- For a support chatbot the duties are the same whether the model runs here or in Virginia: your staff need AI literacy, and from August 2026 a customer must be able to tell they are talking to an AI. Self-hosting is not a way around that, and we do not sell it as one.
This is a description of how the hosting works, not legal advice. Your DPO should read it before you rely on it.
Next
See what a model would do for your work
Every use case names the model we would reach for, what independent tests say about it, and what the hardware would cost.