Model API · served now
Qwen3.8 27B through an OpenAI-compatible API, served from hardware AxForge owns in Sweden. Streaming, tool calls and usage reporting — with zero prompt retention.
Quickstart
Keep the OpenAI client you already use: change base_url,
set your AxForge key and pick the model qwen3.8-27b-nvfp4. Streaming,
tool calls and usage reporting follow the OpenAI API shape.
$ export OPENAI_BASE_URL=https://api.axforge.ai/v1 $ curl -sS "$OPENAI_BASE_URL/chat/completions" \ -H "Authorization: Bearer $AXFORGE_API_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"qwen3.8-27b-nvfp4","stream":true, "messages":[{"role":"user","content":"Summarize DORA in one paragraph."}]}'
# pip install openai from openai import OpenAI client = OpenAI( base_url="https://api.axforge.ai/v1", api_key="YOUR_AXFORGE_KEY", ) r = client.chat.completions.create( model="qwen3.8-27b-nvfp4", messages=[{"role": "user", "content": "Hello from Stockholm"}], ) print(r.choices[0].message.content, r.usage)
Specifications
| Model | Qwen3.8 27B — dense 27B, latest Qwen generation |
|---|---|
| API model name | qwen3.8-27b-nvfp4 |
| Context window | 131,072 tokens |
| Quantization | NVFP4 |
| Capabilities | Chat completions · streaming · tool calls · reasoning output · usage in both modes |
| Hardware | NVIDIA DGX Spark (GB10, 128 GB unified memory) — owned and operated by AxForge |
| Region | eu-se-1 · Stockholm, Sweden — inference stays in-region |
| Input price | €0.29 / 1M tokens |
| Output price | €1.77 / 1M tokens |
Our pricing rule: we target 80% of the lowest current price from a proper inference supplier for the same model, tracked continuously against the market. Found better pricing at a proper inference supplier? Tell us and we’ll look into lowering ours.
Performance
| Metric | Value | Condition |
|---|---|---|
| Decode speed (community) | 25.1 tokens/s | SparkBench PBM @ 4k context, single GB10, NVFP4 + MTP |
| Concurrency | 4 sequences | our serving configuration on one node |
Community measurement from SparkBench for this model on the same hardware class — we haven't published our own tuned numbers yet. More community data on the GB10 page; for quality benchmarks, see the official model card.
Fit
| Use case | Why it fits |
|---|---|
| Private assistants & copilots | Internal or customer-facing assistants with prompts and completions processed in the EU. |
| Tool-using agents | OpenAI-compatible tool calls connect the model to your applications, workflows and data. |
| Long-context analysis | Up to 131,072 tokens for document analysis, synthesis and multi-step tasks. |
| Streaming applications | Tokens stream as they are generated — responsive chat and interactive products. |
| Development & evaluation | A transparent model name, pinned region and usage reporting — prototype before moving to dedicated capacity. |
| EU-hosted inference requirements | Inference runs on hardware AxForge owns and operates in Sweden; TLS terminates in the EU. |
Data & privacy
This describes the customer /v1 API: prompts and completions are
processed in memory in Sweden — not written to disk, not logged, not retained,
and never used to train anything. We keep only request metadata (token counts,
timestamps, status) for billing and operations.
The full policy is at axforge.ai/privacy.
Scale up
Start with the Qwen API — no subscription, no minimum. When you need isolated capacity or control over the machine itself, move to a dedicated NVIDIA DGX Spark for €495/month: the same hardware the API runs on, single-tenant, in the same Stockholm region. Nothing forces the move — the API scales on its own.
FAQ
Yes. Chat completions, streaming, tool calls and usage reporting follow the
OpenAI API shape, so the official OpenAI SDKs work by changing
base_url. The model name is qwen3.8-27b-nvfp4.
In Stockholm, Sweden (region eu-se-1), on NVIDIA DGX Spark
systems AxForge owns and operates. TLS terminates in the EU. Spain
(eu-es-1) is live for the platform; more EU regions are in
deployment.
Launch pricing is €0.29 per million input tokens and €1.77 per million output tokens, excluding VAT. No subscription, no minimum.
131,072 tokens in the production serving configuration — enough for long documents, large tool outputs and multi-step agent work in a single request.
No. Prompts and completions are processed in memory and never persisted or used for training — see the privacy policy. This makes the API a fit for GDPR-sensitive workloads.
On the community SparkBench leaderboard (PBM methodology, 4k context, single GB10), Qwen3.8-27B with NVFP4 + MTP measures 25.1 tokens/s single-stream. We haven't published our own tuned numbers yet.
Yes — a dedicated DGX Spark running Qwen3.8 27B for only your traffic is €495/month on monthly rental. Talk to an engineer to scope it.