Migration
Keep your SDK and your code — change base_url. Open-weight models served from EU regions on hardware AxForge owns, with zero prompt retention. And an honest section on when OpenAI is still the better choice.
The switch
# pip install openai — the official SDK, unchanged from openai import OpenAI client = OpenAI( base_url="https://api.axforge.ai/v1", # was: api.openai.com api_key="YOUR_AXFORGE_KEY", default_headers={"x-axforge-region": "eu-se-1"}, ) r = client.chat.completions.create( model="qwen3.8-27b-nvfp4", # was: a GPT model name messages=[{"role": "user", "content": "Hello from Stockholm"}], ) print(r.choices[0].message.content, r.usage)
$ export OPENAI_BASE_URL=https://api.axforge.ai/v1 $ curl -sS "$OPENAI_BASE_URL/chat/completions" \ -H "Authorization: Bearer $AXFORGE_API_KEY" \ -H "x-axforge-region: eu-se-1" \ -H "Content-Type: application/json" \ -d '{"model":"qwen3.8-27b-nvfp4","stream":true, "messages":[{"role":"user","content":"Summarize DORA in one paragraph."}]}'
Compatibility
| Surface | On AxForge |
|---|---|
| Chat completions | Same API shape — Qwen3.8 27B, 65,536-token context |
| Streaming | Same, with usage reporting in streaming and non-streaming modes |
| Tool calls | Same shape |
| Embeddings | Same shape — Qwen3 Embedding, 1024-dim vectors |
| Images | Generation endpoint — ERNIE Image Turbo, €0.02 / image; editing via FLUX.2 Klein |
| Audio | Transcription (Whisper) and speech (Piper) |
No spin
| Difference | The honest version |
|---|---|
| Models | Open-weight models, not GPT. Qwen3.8 27B is served now; further open-weight models are deployable on request on a dedicated system. Run your own evals — the switch is one line of code. |
| Model menu | Smaller and curated. OpenAI's catalogue and feature surface are far larger. |
| Hosting | EU regions on named hardware: NVIDIA DGX Spark (GB10) systems AxForge owns in Sweden. TLS 1.3 terminates in the EU. You can pin and verify the region on every call. |
| Data handling | Zero prompt retention — prompts and completions processed in memory, never written to disk, logged, retained or used to train. Metadata only (token counts, timestamps, status). |
| Pricing | Published: Qwen3.8 27B at €0.29 / 1M input · €1.77 / 1M output (launch pricing); image generation €0.02 / image; the rest is on the price list. Priced at 80% of the lowest market rate for the same model — run it against your current bill. |
| Access | API keys are allocated from a waiting list today — not instant self-serve. |
Credibility
If your workload needs frontier-model quality — the hardest reasoning, coding or multimodal tasks — OpenAI's latest models will beat a 27B open-weight model, and pretending otherwise would waste your evaluation time. The same goes if you depend on OpenAI-specific features, its much larger model menu and ecosystem, or instant self-serve scale. In those cases, stay.
Choose AxForge when the binding constraints are different: prompts that must be processed in the EU, a retention posture you can defend in writing, and knowing exactly which machine serves you. That is what we build — GDPR compliant EU infrastructure — and Qwen3.8 27B handles a wide range of production chat, extraction and tool-use workloads at a published price.
Data & privacy
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.
FAQ
Yes. Chat completions, streaming, tool calls and usage reporting follow the
OpenAI API shape, so the official SDKs work by changing base_url
to https://api.axforge.ai/v1 and swapping the model name.
There are no GPT models here. The served chat model today is
Qwen3.8 27B
(qwen3.8-27b-nvfp4), an open-weight model; further open-weight
models are deployable on request on a dedicated system. Most teams find the
switch takes an afternoon — run your own evals and see.
Our prices are published and aggressively low — Qwen3.8 27B at €0.29 in / €1.77 out per million tokens — priced at 80% of the lowest market rate for the same model. Run them against your current bill.
Frontier-model quality on the hardest tasks, a much larger model menu and ecosystem, and instant self-serve signup — our keys come from a waiting list today. What you gain: EU hosting on named hardware, zero prompt retention, and transparent serving.
EU regions on hardware AxForge owns: eu-se-1 (Stockholm) and
eu-es-1 (Málaga) are live, more EU regions in deployment.
Inference runs on NVIDIA DGX Spark (GB10) systems in Sweden. Prompts are
processed in memory — never stored, logged or used to train; only billing
metadata is kept. See the privacy policy.
When frontier quality is the binding constraint, when you depend on OpenAI-specific features or ecosystem, or when you need instant self-serve scale. If EU residency, zero retention and hardware transparency matter more, that is what we are built for.
Related