Qwen3.8-27B API
Hosted inference for Qwen3.8-27B, served from EU hardware in NVFP4 on Blackwell. OpenAI-compatible endpoint, streaming, prompt caching, 262,144-token context. The performance numbers below were measured on the card in service on 30 September 2026, with the decision models loaded and idle. What live traffic looks like right now is on the status page.
| Input | Output | Cached input |
|---|---|---|
| $0.25 | $2.09 | $0.04 |
Prompt prefixes served from cache bill at the cached-input rate, about one sixth of the input price. No subscription, no minimums. Volume pricing for sustained load: artem@llmtech.eu.
from openai import OpenAI
client = OpenAI(
base_url="https://api.llmtech.eu/v1",
api_key="YOUR_KEY", # get one: artem@llmtech.eu
)
r = client.chat.completions.create(
model="nvidia/Qwen3.8-27B-NVFP4",
messages=[{"role": "user", "content": "Hello"}],
stream=True,
)
for chunk in r:
print(chunk.choices[0].delta.content or "", end="")
API keys are issued by hand while we are in early production — email artem@llmtech.eu and you will have one the same day. Self-service is planned.
Zero data retention
Prompts and completions are processed in volatile memory only — never written to disk, logs, analytics or backups, and never used for training. We retain per-request technical metadata only: timestamp, request ids, key fingerprint, pricing channel, source IP address, model, stream flag, requested reasoning mode, token counts, prices and cost, HTTP status, error code and durations.
GDPR compliance and DPA
LLM Tech operates from the EU (Poland) and processes all data on EU hardware. A signable Data Processing Agreement under Art. 28 GDPR is available on request, including our technical and organisational measures and the full sub-processor list (infrastructure only — no sub-processor has access to request content).
Where does inference run?
All compute is located in the European Union: the GPU node is in Italy, and the edge is in Nuremberg, Germany. No data leaves the EU/EEA and there is no US cloud at any hop. TLS 1.3 to our edge, then an encrypted WireGuard tunnel to the GPU node.
How fast is it, really?
Time to first token is dominated by prompt length, so a single number would be misleading. On a short prompt we measure 0.13 s, and a 36,000-token prompt already in cache comes back about nine times faster than the same prompt cold. Generation runs at 125 to 144 tok/s when a stream has the card to itself, 85.6 tok/s with twelve streams running and 47.0 tok/s with thirty-two; fresh prefill tops out at 11,100 tok/s. Those figures were measured on 30 September 2026 on the card in service, with the decision models loaded and idle. We would rather publish the spread than quote our best case: the status page shows the current numbers, whatever they are today.
How do I verify your performance claims?
The status page publishes TTFT and generation speed measured on client traffic, and uptime from a probe we send every minute, refreshed every 5 minutes, with an hourly strip for the last 24 hours. That is a different measurement from the figures above and usually a lower number, because a card serving customers is never idle.
Does the model accept images?
Yes. The NVFP4 quantisation keeps the vision tower, so image input works. Image tokens are counted into prompt_tokens and itemised at usage.prompt_tokens_details.multimodal_tokens, billed at the input rate — a 64×64 PNG costs 64 image tokens. Most third-party listings for this model describe it as text-only; that is wrong. Images must go in a user message, not a system message.
How does prompt caching work, and what does it cost?
Repeated prompt prefixes are cached automatically and billed at $0.04 per million instead of $0.25. The cache uses 1,600-token blocks and reuses every whole block except the last one, so a prompt needs at least 3,200 tokens to get any cache; a repeated prefix is read from cache from its second request on. Keep the system prompt and tool definitions byte-identical and in a stable order, and put anything volatile after them. On our own production traffic this brings the effective input price to about $0.07 per million.