// open models · european infrastructure
DeepSeek, GLM and Qwen,
served from Europe
The best open models, served on our infrastructure without DeepSeek, Z.ai or Alibaba seeing your data.
- Own hardware in the EU
- No prompt logs
- GDPR DPA
- OpenAI-compatible
// where your data is processed
Open models running
on our machines
DeepSeek, Z.ai and Alibaba publish the weights of these models and we run them on our own servers. The weights were downloaded once, when the model was deployed, and the developers receive no prompts, no responses and no usage data.
// request lifecycle
What happens
to a request
The request enters our infrastructure, is processed in memory and the response returns to your application. We do not keep the content of the prompt or the response once inference ends.
// models
The models we operate
in Europe
Open weights published by their developers, served from the same Helmcode API.
| Model | Developer | Served by | Runs in | Type | Context | Licence | Endpoint |
|---|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | DeepSeek · China | Helmcode | Europe | LLM | 1M | MIT | deepseek-v4-flash |
| GLM 5.3 | Z.ai · China | Helmcode | Europe | multimodal LLM | 1M | GLM-5.3 License* | glm5.3 |
| GLM 5.3 Flash | Z.ai · China | Helmcode | Europe | multimodal LLM | 1M | MIT | glm5.3-flash |
| Qwen 3.6 | Alibaba · China | Helmcode | Europe | multimodal LLM | 256K | Apache 2.0 | qwen3.6 |
| Qwen3 Embedding 8B | Alibaba · China | Helmcode | Europe | embeddings | · | Apache 2.0 | qwen3-embedding |
| Qwen3 Reranker 8B | Alibaba · China | Helmcode | Europe | reranking | · | Apache 2.0 | rerank |
* GLM-5.3 License: MIT plus a Z.ai clause for anyone earning over $10bn a year serving the model.
// privacy
Your data stays
safe and private
Every request is processed in memory, in the EU, and discarded when inference ends. This is everything that happens to what you send to the open models we operate.
Processed · transient
Only while inference runs
The prompt and the response live in GPU memory for the length of the request, inside the EU.
- Your prompt, in memory, in the EU
- Inference on the model
- The response, streamed back to you
Stored · for billing only
Just enough to bill you
Metadata and counts that let us operate and invoice the service. Never the content.
- API key metadata
- Aggregate request and token counts
Never stored
Nothing you send stays
No prompt or response is written to disk, and nothing you send trains a model.
- Prompts
- Responses
- Your documents or code
- Anything that trains a model
// cost
The difference shows
on the invoice too
Example for 10 billion tokens a month, 80% input and 20% output.
-
Claude Opus 5.5 via API
list price
~$72,000/ month
-
DeepSeek V4.1 Flash via the developer’s API
off-peak price; it doubles at peak hours
~$2,400/ month
-
Helmcode Growth
includes 15B tokens a month of DeepSeek V4.1 Flash and GLM 5.3 Flash; Qwen 3.6 unlimited
€1,299/ month
The comparison shows the cost of processing that volume of tokens, not an equivalence of quality or capability between the models.
Official prices in October 2026. Bars on a linear scale of each figure.
see_pricing →// security and procurement
What your security team
will ask you
These answers cover the open models Helmcode operates. Copy them as they are, or request the DPA and the full pack.
- 01
Is personal data transferred outside the EEA during inference?
No, for the open models Helmcode operates: inference runs on Helmcode infrastructure in EU data centres. Closed models resold through the same API run on their providers’ infrastructure, under their terms.
- 02
Are DeepSeek, Z.ai or Alibaba subprocessors?
No. They publish the model weights. Helmcode downloads them and runs them on its own infrastructure; the developers receive no requests and no customer data.
- 03
Are prompts or responses stored or logged?
No. Prompts and responses are processed in memory and discarded when inference ends. Helmcode keeps only API key metadata and aggregate request and token counts, for operating and billing the service.
- 04
Is customer data used to train or fine-tune models?
No. Nothing sent to the API is used to train or fine-tune any model.
- 05
Can the model send data to a third party by itself?
No. The weights are loaded into the inference engine Helmcode operates. The model has no network connection of its own to DeepSeek, Z.ai, Alibaba or anyone else.
- 06
Which jurisdiction does the service operate under?
The EU. These models run on infrastructure Helmcode owns and operates in the EU, and Helmcode signs a Data Processing Agreement under Article 28 of the GDPR.
// faq
Details that usually
come up next
Is it exactly the same model the developer offers?
It is the open-weights release the developer published. Some models are served quantised to NVFP4 for efficient inference, so answers can differ slightly from the developer’s endpoint. Test it on your own cases before production.
Where are the servers?
In EU data centres, on hardware Helmcode owns and operates.
What about the closed models available on Helmcode?
They run on their providers’ infrastructure, under their terms. Everything on this page about EU processing and no international transfer refers only to the open models Helmcode operates.
Can I deploy it on-premise?
Yes. The on-premise deployment runs the same models and the same API inside your own infrastructure.
Try these models
on your own workload
The same OpenAI-compatible API. Inference in the EU, no prompt logs and a flat rate.