Unlimited tokens
No caps on consumption — only RPM and concurrency per API key.
// product
Run open frontier models
through a single API.
Open models, operated for you on EU infrastructure, the full inference stack behind a single endpoint. Sovereignty + flat rate + zero logs.
// architecture
One endpoint. It never leaves the EU.
A request hits a single OpenAI-compatible URL, is routed and rate-limited in our control plane, and answered by open models on managed GPUs — all inside the EU, none of it logged.
Your request enters one EU endpoint and never leaves — no prompt is stored, no data crosses to a US hyperscaler.
// guarantees
Four things, by architecture.
Not features you configure — properties of how the platform is built. They hold for every model and every use case.
OpenAI-compatible
Change the base URL and key. Every OpenAI-compatible client works as-is.
Zero logs
Prompts are never stored. Your data and code never train a model.
Data in the EU
Processed only on EU infrastructure — not subject to the Cloud Act.
// the platform
Explore the stack.
Four areas, one product. Go as deep as you need on models, where it runs, how it's secured and what it plugs into.
Models
Nine open models — LLMs, embeddings, reranking and speech — behind one API.
Deployment
Shared, dedicated GPU or fully on-premise — same stack, different sovereignty.
Security & Compliance
Zero logs, EU residency, AI Act native, GDPR and DORA by architecture.
Integrations
Cursor, Zed, OpenCode, LangChain, the OpenAI SDK — drop-in, unchanged.
// capabilities
Everything the API can do.
One OpenAI-compatible endpoint, the full feature surface — text, vision, voice, retrieval and agents.
Tool & function calling
Native function calling with the OpenAI JSON schema — agents that act, not just chat.
all LLMs Structured outputs
Constrain responses to your JSON schema with response_format — typed, every time.
response_format Vision & multimodal
Image and audio input on Gemma 4 and MiMo — read scans, charts and screenshots.
gemma4 · mimo Streaming
Token streaming over SSE for real-time chat, copilots and voice UX.
SSE Long context
Up to a 1M-token context window on DeepSeek V4-Flash — whole corpora in one pass.
up to 1M Embeddings & reranking
4096-dim multilingual vectors plus cross-lingual reranking — retrieval, built in.
qwen3-embedding · rerank Speech · STT & TTS
Whisper transcription and Kokoro synthesis — 99+ languages, sub-second voice.
whisper · kokoro Unlimited tokens
No caps on consumption — limits are RPM and concurrency per API key.
per API key // by the numbers
The platform, spec'd out.
The hard numbers behind the stack — context, hardware, region and reliability.
Context window up to 1M tokens
Embedding dims 4096
Models 9 in production
Hardware B200 · 192GB
Region EU · Madrid
Uptime SLA 99.9%
API OpenAI-compatible · 6 endpoints
Data retention zero logs
// use cases
What teams build on it.
The same stack powers retrieval, voice, copilots, document workflows and agents — each with its own playbook.
// product faq
The platform, answered.
What teams ask before moving inference onto Helmcode.
What does Helmcode actually run?
Open-weight models — DeepSeek, Qwen, Gemma, plus embeddings, reranking and speech — served behind an OpenAI-compatible API and operated by us on EU GPUs, with zero logs.
How do I get started?
Get an API key from the console, change your base URL and key, and you're running. Any OpenAI-compatible SDK or tool works unchanged — most teams ship the same day.
Which models are available?
Nine in production: DeepSeek V4-Flash, MiMo, Qwen 3.6 and Gemma 4 for text, qwen3-embedding and rerank for retrieval, and Whisper and Kokoro for speech. See the Models page for specs.
Where does inference run?
Exclusively on EU infrastructure — never on US hyperscalers subject to the Cloud Act. GDPR and AI Act native, by architecture rather than configuration.
Is it managed, or do I self-host?
Fully managed: we provision, monitor and operate the whole stack. For stricter needs you can move to dedicated GPUs or a full on-premise deployment inside your own datacenter.
How is it priced?
Per API key — a flat monthly rate, not per token. Unlimited tokens on open models, no usage surprises, no lock-in. See Pricing for plans.
// get started
START BURNING TOKENS
Skip the AI infra work. Deploy your first private inference endpoint today.
Flat rate. EU data. OpenAI API compatible.