Sovereign AI inference.
Your data never leaves Europe.
Open models on dedicated GPUs inside the European Economic Area. Zero data retention. OpenAI- and Anthropic-compatible. Built for agents.
EU company · no US parent · prompts never logged
from openai import OpenAIclient = OpenAI(api_key="hb-...",base_url="https://api.heabsy.com/v1",)
One changed line. Streaming, tool calling and structured output work as before.
per stream — about 2× the fastest provider of this model on the OpenRouter showcase
time to first token on free capacity; 0.9 s under load
prefix-cache hit rate on agentic workloads, even at the edge of the 131k window
faster answer on a repeated 34k-token prompt: 0.48 s versus 42 s
Measured on the live system · Aug 2026 — nothing on this page is a projection.
Running AI in production in Europe has three doors today. Each one costs you something.
US gateways and inference clouds
Fast and cheap, but your prompts cross the Atlantic and land under the CLOUD Act — wherever the servers sit, a US company answers to US law. For a bank, a hospital or a public agency that is not a discount, it is a disqualifier.
Hyperscalers, EU region
An EU region of a US hyperscaler is still a US company — and running open models on rented GPUs yourself means quotas, capacity planning and an inference stack you now operate.
Do it yourself
Renting GPUs and serving a model is a weekend project. Prefix caching, cache-aware routing, priorities under saturation and honest usage accounting are not. That is the 80% of the work that starts after the model loads.
Nobody offers production-grade inference on European hardware, under a European company, with prompts that are never stored. That gap is where Heabsy runs.
Our GPUs. Our stack. EU jurisdiction.
The flagship model runs on machines we operate in Italy and Norway — both inside the European Economic Area. For it we own the serving stack end to end, which is why we can promise what happens to your data and prove how fast it moves.
The operating company is incorporated in the EU with no US parent — the CLOUD Act does not reach it. Jurisdiction is a property of the company, not of the server room.
EEA-only processingEU companyOwn hardwareZero data retention
What we never do with your data
- Prompts are never written to logs
- Completions are never written to logs
- Nothing is used to train any model
- No request content leaves the EEA
Usage accounting stores token counters and metadata — never the text. In your dashboard you see fresh, cached and output tokens per request; we see the same counters and nothing else.
Roughly twice the fastest public provider of the same model.
~176 tokens per second on a single stream, first token in 0.3–0.9 s. The gain comes from a trained speculative-decoding draft (MTP, depth 3) running on dedicated RTX 5090s — not from serving a smaller model than advertised.
Long context slows any GPU — that is memory bandwidth, physics rather than queueing. At 50–110k of context we still decode at 79–95 tok/s, and the first token keeps arriving in under a second because of the prefix cache below.
Don't take our word for it. Get a key and measure on your own workload — the usage report shows latency and token counts for every request you send.
Tokens per second, single stream. Showcase figures: openrouter.ai, providers of the same model, Aug 2026.
Agents reread. We stopped charging compute for it.
On agentic workloads 98.5% of all processed tokens are context the model has already read. Our router keeps a conversation on the machine that holds its prefix, so that context is never recomputed — the first token stays under a second even at 100k of context.
Measured on the live system: a repeated request with a 34,000-token prompt is answered in 0.48 seconds from the prefix cache, versus 42 seconds cold — 84 times faster.
Replayed at ~12× speed. Measured on the live system · Aug 2026.
of tokens in agentic coding sessions are rereads of known context — served from cache, not recomputed
time to first token at 98k tokens of context on a warm conversation
cache hit rate at the very edge of the 131k window
This is why Claude Code, Cursor and Cline feel different here: an agent that iterates ten times pays the full read once, and the other nine are light. The cached_tokens field is reported on every response, so you can audit the saving instead of trusting us.
Our own model — and the catalog around it.
The flagship runs on our hardware in the EEA — everything measured above was measured on it. Around it, a routed catalog of open models: bought from global providers, sold by a European company on one invoice, behind the same key.
Qwen3.8 27B
our own hardwareNo minimum spend, no subscription. The fleet is elastic — new GPUs register automatically and capacity scales roughly within an hour. Live machine-readable catalog: api.heabsy.com/openrouter/v1/models.
Routed open models, same key
Fulfilled through global providers under our European contract — one DPA, one invoice, an EU counterparty. Compute may run outside the EEA; for strict EEA-only processing, use the flagship above.
GLM 5.3 Flash
gpt-oss-120b
DeepSeek V4 Flash 0423
Hy3
MiMo-V2.5
Mistral Small 4
Prices from the gateway price list, synced 2026-08-29. The bill follows the gateway, and this page follows the gateway.
Keep your SDK. Change one URL.
The API is OpenAI-compatible — chat completions, SSE streaming with usage reporting, tool calling and structured output work with any OpenAI SDK. The Anthropic-compatible endpoint is not a promise either: this platform serves as a Claude Code backend in daily use.
- Per-key budgets and rate limits (RPM / TPM)
- Usage split into fresh, cached and output tokens per request
- Early 429s under saturation instead of a silent queue
Three steps, no credit card.
- 01Write to us — we create your account and set a spending limit.
- 02Sign in to the console and issue as many keys as you need, each with its own budget.
- 03Point your client at the endpoint. No code changes — OpenAI- and Anthropic-compatible.