Qwen3.8-27B API from EU hardware. Measured, not promised.
curl https://api.llmtech.eu/v1/chat/completions \
-H "Authorization: Bearer $LLMTECH_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "unsloth/Qwen3.8-27B-NVFP4",
"messages": [{"role": "user", "content": "Hello"}],
"stream": true
}'
OpenAI-compatible: point your existing SDK at https://api.llmtech.eu/v1 and it works. Free trial key in the docs — no signup, no card.
| Provider | Input, $/M | Output, $/M | Cached input, $/M | Context |
|---|---|---|---|---|
| LLM Tech | $0.25 | $2.09 | $0.04 | 262,144 |
| Chutes | $0.35 | $2.75 | $0.035 | 262,144 |
| DeepInfra | $0.40 | $3.00 | $0.04 | 262,144 |
| Alibaba Cloud | $0.50 | $3.00 | $0.10 | 1,000,000 |
prices checked 2026-08-25 on public pricing pages and marketplace listings; tell us if we missed a cheaper one — we intend to stay the lowest.
Zero data retention
Prompts and completions live in volatile memory only. Never written to disk, logs, analytics or backups. Never used for training.
EU jurisdiction and hardware
Polish company. German edge (Hetzner), Finnish GPU datacenter (Verda). No US cloud anywhere in the stack.
GDPR-ready
Signable Art. 28 DPA, published sub-processor list, documented technical measures. Security page →
How do I get an API key?
Try the free trial key from the docs right now. For a production key, email artem@llmtech.eu — issued the same day, usually within the hour. Self-service is planned.
How does billing work?
Per token, monthly invoice: $0.25/M input, $2.09/M output, $0.04/M cached input. No subscription, no minimums. Every request is journaled with its price snapshot, so an invoice is verifiable line by line.
Are your performance numbers real?
They are measured on live production traffic and published at llmtech.eu/status, refreshed every 5 minutes, including an hourly uptime strip. Nothing on this site is quoted from a datasheet.
Can I control the model's reasoning?
Yes: chat_template_kwargs.enable_thinking (on/off) and reasoning_effort (low / medium / xhigh). Thinking is adaptive by default — the model skips it on trivial prompts.
What about rate limits and SLA?
64 concurrent requests per key by default, higher on request. Above the limit you get an instant 429 with Retry-After — requests never queue silently. Uptime history is public on the status page.