Trust centre · Transparency
Model transparency
You always know which model answered, how it is served, and who made it. This page is the policy; the model pages carry the specifics for every model we serve.
The policy
What we publish for every served model
| Commitment | Where you see it |
|---|---|
| Served model identity | Returned on every API response — you never have to guess which model produced an output. |
| Quantization | Disclosed in the specs on each model page (written as NVFP4 where used). Never hidden. |
| Developer & original model card | Each model page names the model's developer and links the original model card. For benchmarks, see the model card — we publish only our own measured numbers. |
| Changes & retirements | Dated notes on the model page — see the change and retirement policies below. |
AI Act
Article 50 position
Article 50 transparency duties of the EU AI Act apply since 2 August 2026. AxForge outputs are machine-generated. Our part and yours:
| Point | Position |
|---|---|
| Machine generation | All AxForge API outputs are machine-generated, and every model page states it. |
| Disclosure support | The API returns the served model identity on every response, giving your product what it needs to disclose machine generation to its users. |
| Customer duty | Customer-facing products built on the API must disclose machine generation to their users. |
| Embedded provenance | No machine-readable content-provenance metadata (e.g. C2PA) is embedded in outputs today. It is planned. |
Change policy
How models change — launch draft
| Event | Policy |
|---|---|
| Additions & replacements | Announced with dated notes on the models pages. |
| Major replacements | Announced 14 days ahead once the service is in production. While the service is pre-release, changes may be faster — always dated on the page. |
| Identity guarantee | The served model identity is always returned in API responses, so any change is visible in your own logs, not only ours. |
Retirement policy
How models leave the shared API
Models leave the shared API only with a dated note on the model page. Dedicated deployments are unaffected — a model retired from the shared API can still run on a dedicated system for as long as you need it.
- Policy in practice: Qwen-Image left the shared API on 2026-08-26 with a dated note on its page; it remains deployable on request on dedicated hardware.
Catalogue
Served models: developers and licences
| Model | Developer | Licence |
|---|---|---|
| Qwen3.8 27B | Alibaba (Qwen team) | Open weights |
| Qwen3 Embedding | Alibaba (Qwen team) | See the model card |
| ERNIE Image Turbo | Baidu | Apache 2.0 |
| FLUX.2 Klein 4B | Black Forest Labs | Apache 2.0 (4B variant) |
| MiniMax Music 3 | MiniMax | Community License |
| Whisper | OpenAI | MIT |
| Piper | rhasspy | MIT |
- The Apache 2.0 licence applies to the FLUX.2 Klein 4B variant we serve; the 9B variant is non-commercial and is not served.
- Full licence terms live in each developer's original model card, linked from the model pages.
- Models deployable on request (not on the shared API) are listed in the model catalogue, marked as such.
Related
Nearby on AxForge
Document control
Version
Version 0.1 (launch draft) · Effective 2026-08-26
- Change log: 2026-08-26 — first published.