Image API · served now
Image generation through the OpenAI-compatible /v1/images/generations endpoint, served from hardware AxForge owns in Sweden. Dense, legible in-image text — posters, infographics and UI-style layouts that actually read — with zero prompt retention.
Specifications
| Model | ERNIE Image Turbo — 8B diffusion transformer, 8-step turbo |
|---|---|
| API model name | ernie-image-turbo |
| Endpoint | POST /v1/images/generations |
| Signature capability | Dense, long-form, layout-sensitive text inside the image — posters, infographics, UI-style images, multi-panel layouts |
| License | Apache 2.0 (open weights) |
| Quantization | fp8 |
| Hardware | NVIDIA DGX Spark (GB10, 128 GB unified memory) — owned and operated by AxForge |
| Region | eu-se-1 · Stockholm, Sweden — pinned with x-axforge-region, echoed on the response |
| Price | €0.02 / image |
Picked in a seed-matched bake-off on our own hardware: ERNIE won the structured and text-heavy cases decisively. Image editing runs on FLUX.2 Klein 4B.
Performance
| Metric | Value | Condition |
|---|---|---|
| Generation time | ~32 s | per 1024×1024 image, 8-step turbo + prompt enhancement |
Measured on our production DGX Spark node, single-stream, 2026-08. For quality benchmarks, see the official model card — we publish only numbers we measured ourselves.
Quickstart
$ export OPENAI_BASE_URL=https://api.axforge.ai/v1 $ curl -sS "$OPENAI_BASE_URL/images/generations" \ -H "Authorization: Bearer $AXFORGE_API_KEY" \ -H "x-axforge-region: eu-se-1" \ -H "Content-Type: application/json" \ -d '{"model":"ernie-image-turbo","size":"1024x1024", "prompt":"Minimal concert poster, bold type reading OPEN MODELS / EU HARDWARE"}'
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
€0.02 per image. No subscription, no minimum.
We measured roughly 32 seconds per 1024×1024 image (8-step turbo with prompt enhancement) on our production DGX Spark node, single-stream, 2026-08. We publish measured numbers only.
Yes. It is the standard /v1/images/generations shape at
https://api.axforge.ai/v1, with the model name
ernie-image-turbo.
Yes — dense, long-form, layout-sensitive text is the model's signature capability: posters, infographics, UI-style images and multi-panel layouts where every word has to read.
We benchmarked candidates on our own hardware with identical prompts and seeds. ERNIE Image Turbo won the structured and text-heavy cases decisively, under the Apache 2.0 license, at a third of the memory of the previous model. Editing runs on FLUX.2 Klein 4B.
No. Prompts and completions are processed in memory in Sweden, never persisted or used for training — see the privacy policy. This makes the API a fit for GDPR-sensitive workloads.