A dense 27-billion-parameter open-weights model, built for reasoning, code and multilingual dialogue. We operate it on our European GPUs and you call it like any OpenAI-compatible API.
Parameters
27 bn
Context
262 K
Languages
100+
Served throughput
~90 t/s
Toggle between fast answers and step-by-step reasoning for mathematics, logic and code — in the same endpoint, with a single parameter.
Structured JSON output and declared tools: the model slots into your existing agents and workflows without an adaptation layer.
Trained on more than 100 languages and dialects: one model serves your customers from Lisbon to Helsinki, with no intermediate translation.
Permissive licence and public weights: you know exactly what is running, and you can replicate the inference in-house the day you choose. Why Mistral and Qwen are champions of European sovereignty → · Cybersecurity: the open-weights answer →
| Benchmark | Qwen3.8-27B | 20-30 bn class |
|---|---|---|
| MMLU-Pro (knowledge) | 72,7 | 65-74 |
| GPQA (science) | 53,5 | 45-56 |
| AIME (mathematics) | 30,3 | 18-34 |
| LiveCodeBench (code) | 28,8 | 22-31 |
Indicative values taken from the model publisher’s publications; definitive figures will be measured on our infrastructure and published on this page.
OpenAI SDK, LangChain, LlamaIndex: everything works by changing the base URL and the key. The reasoning parameter switches on reasoning mode.
from openai import OpenAI
client = OpenAI(
base_url="https://api.europeangpu.com/v1",
api_key=EGPU_API_KEY)
r = client.chat.completions.create(
model="qwen3.8-27b",
messages=[{"role": "user",
"content": "Summarise this contract"}],
extra_body={"reasoning": True})No subscription, no minimum — the price of sovereign inference, with no hidden cost. Need more throughput for less? See Ministral 3 14B →