Serving DeepSeek V4 Flash · Early access open
Frontier AI inference.
Powered by Iceland.
icetok runs open frontier models on 100% renewable energy — with token speeds that rival the big clouds and a guarantee, by architecture, that your data never leaves the country.
- OpenAI-compatible API
- 1M token context
- EU/EEA-owned
- EU data residency
401 tok/s
Single-stream output · 2K context
1,863 tok/s
Aggregate throughput · 24 concurrent
1M
Token context window
100%
Renewable energy, always
For developers
Point your SDK at icetok. Nothing else changes.
A drop-in, OpenAI-compatible API served from Iceland. Same client libraries, same schemas — frontier open models with serious throughput and data residency baked in.
# Frontier model. Iceland. One call.
$ curl https://api.icetok.is/v1/chat/completions \
-H "Authorization: Bearer $ICETOK_API_KEY" \
-d '{
"model": "deepseek-v4-flash-0731",
"messages": [
{"role": "user", "content": "Why is the sky green here?"}
],
"stream": true
}'
-
OpenAI-compatible API
Chat Completions, streaming, tool calls and embeddings. Swap the base URL, keep your stack.
-
1M token context
Feed an entire codebase or document corpus in one request — ingest at ~1,000 tok/s prefill.
-
Fast by default
Speculative decoding, prefix caching and chunked prefill out of the box. No tuning to start.
-
Open models, no lock-in
Open-weight models with permissive licenses. Bring your own open model — we’ll serve it. Nothing stops you from deploying elsewhere.
-
MCP-ready
Point agents and office tools at your icetok endpoint over MCP — your tools, your keys, your data path. No proprietary agent stack required.
Benchmarks
The numbers that made us build this
Measured on reproducible public benchmarks of DeepSeek V4 Flash 0731 (284B MoE) on a single 2× RTX PRO 6000 Blackwell 96 GB node — the same class of hardware icetok runs.
| Concurrent requests | 2K context | 100K context |
|---|---|---|
| 1 | 401 tok/s | 247 tok/s |
| 4 | 790 tok/s | 405 tok/s |
| 8 | 1,263 tok/s | — |
| 24 | 1,863 tok/s aggregate | — |
TP=2, DSpark speculative decoding, NVFP4 KV cache, prefix caching, CUDA graphs. Throughput varies with prompt mix and context — these are the published headline figures, August 2026.
For the public sector
Sovereign AI. By architecture, not by contract.
For governments, agencies and enterprises, the question isn’t just how fast — it’s where. Every token is generated in Iceland. icetok is built so your data physically cannot leave the country.
Router
│
┌────────────┼────────────┐
▼ ▼ ▼
Node A Node B Node C
2× PRO 2× PRO 2× PRO
6000 6000 6000
~0.4–1.8k ~0.4–1.8k ~0.4–1.8k
tok/s agg. tok/s agg. tok/s agg.
│
X
NO US CLOUD
NO API EGRESS
NO TRAINING ON DATA
-
Owned in Europe
An EU/EEA-owned company under Icelandic jurisdiction, not a US subsidiary renting an EU region. Under US law (FISA 702, CLOUD Act), data held by a US company can be compelled — even from its “EU” data centers. Sovereignty is ownership, jurisdiction and architecture, not a region label.
-
Data stays in Iceland
Inference runs on our own nodes in Icelandic colocation facilities, on an open-source serving stack. Redundant replica pairs — one can fail and the service continues.
-
No training on your data
Your prompts are never logged for training, never shipped abroad, never used to improve third-party models. A guarantee, not a promise.
-
GDPR & compliance pack
Processing in the EEA under GDPR. DPA-included, ISO 27001 (in certification), EUCS "High" on the roadmap — built for the EU's strictest procurement requirements.
-
100% renewable power
Geothermal and hydroelectric energy, one of the lowest carbon intensities in Europe, free cooling year-round.
How it works
From key to first token in minutes
-
01
Request early access
Leave your email below. Founding members get keys first and a seat at launch pricing.
-
02
Get your API key
We provision your key and point you at the endpoint. OpenAI-compatible, so your existing code just works.
-
03
Scale on sovereignty
Start with shared throughput, move to a dedicated replica pair when you’re ready. One router, your own nodes.
FAQ
Questions, answered
Which models does icetok run?
Today we serve DeepSeek V4 Flash 0731 — a 284B-parameter open MoE model with a 1M token context window. We also serve other permissively-licensed open models, and can host models you bring.
Is the API really OpenAI-compatible?
Yes. Chat Completions, streaming, tool calling and embeddings use the same schemas. Point your existing OpenAI SDK at our base URL and change the key. No migration work.
Where is my data processed?
Exclusively in Iceland, on our own nodes in Icelandic colocation facilities. There is no US cloud hop, no API egress, and no data transfer for training. This is an architectural property of the service, not a promise in a contract.
How is this different from an EU region of a US cloud?
An EU region of a US provider is still a US company. Under US law (FISA 702, the CLOUD Act), US companies can be compelled to disclose data held anywhere — including in their European data centers. icetok is an EU/EEA-owned company operating its own nodes in Iceland: no US parent, no US jurisdiction, no US cloud hop. The difference is structural, not a label.
What certifications and compliance documents do you provide?
At launch, every contract ships with a signed DPA and a transparent subprocessor list (default: none for inference). ISO 27001 is in certification; EUCS "High" — the coming EU cloud security scheme — is on the roadmap. Target service availability is 99.9%, with SLAs published before production launch.
Do you train on my prompts or data?
Never. Prompts and generated content are not logged for training, not retained for model improvement, and not shared with any third party. Dedicated customers can additionally sign a DPA with full deletion guarantees.
How is icetok sustainable?
Iceland's grid is nearly 100% renewable — geothermal and hydroelectric. Cool Nordic air provides free datacenter cooling, and our hardware runs at a fraction of hyperscale power draw. Per-token carbon is among the lowest you'll find anywhere.
What does it cost?
Per-token pricing at launch, published transparently — no reserved-instance games. Early access members lock in founding rates. For public sector, we offer procurement-ready terms, DPAs and dedicated capacity.
Can we get a dedicated, private deployment?
Yes. Dedicated replica pairs in Icelandic colocation, isolated from shared traffic, with your own capacity and routing. Reach out through early access and we’ll scope it with you.
Be first on the fast, sovereign inference cloud.
Early access is open. Founding members get API keys first, founding pricing, and a direct line to the team.
Built in Iceland 🇮🇸 · hello@icetok.is