Nebius

F Nebius Token Factory (formerly AI Studio) inference, fine-tuning and Sandboxes API

Token Factory's OpenAI-compatible inference, fine-tuning, batch and dedicated-endpoint APIs take a static project API key created by hand in the web UI; no expiry, rotation API, last-used view or GitHub partner revocation is documented. Only the Sandboxes API also takes a Nebius IAM token. Fine-tuning pushes to Hugging Face and W&B with tokens you paste in.

Docs: docs.tokenfactory.nebius.com · Checked 2026-10-22 · source YAML · dispute via PR

Inbound F

How a workload running anywhere authenticates to this product.

Static API key created on the project's API keys page in the Token Factory UI and sent as Authorization: Bearer (NEBIUS_API_KEY) to the inference, fine-tuning, files, batch, datasets and dedicated-endpoint APIs and to Prometheus metrics scraping. The Sandboxes API alone documents a Nebius IAM bearer token plus Project header as its recommended scheme.

docs.tokenfactory.nebius.com checked 2026-10-22
“In Nebius Token Factory, go to the [API keys](https://tokenfactory.nebius.com/project/api-keys) section.”
docs.tokenfactory.nebius.com checked 2026-10-22
“Keep your keys private; do not share or expose them in client-side code.”
docs.tokenfactory.nebius.com checked 2026-10-22
“api_key=os.environ.get("NEBIUS_API_KEY")”
docs.tokenfactory.nebius.com checked 2026-10-22
“scheme: bearer”
docs.tokenfactory.nebius.com checked 2026-10-22
“Issue a [static API key in Nebius Token Factory UI](https://tokenfactory.nebius.com/project/api-keys).”
docs.tokenfactory.nebius.com checked 2026-10-22
“This is the recommended authentication scheme for new clients.”
docs.tokenfactory.nebius.com checked 2026-10-22
“Supports two auth types: iam (default) — bearer token + project ID, default URL provided jwt (legacy) — bearer token only, URL must be specified”

Graded on the inference and fine-tuning API, the path every workload must use. Its OpenAPI security scheme is a plain HTTP bearer and every guide uses NEBIUS_API_KEY; no Nebius IAM token, service-account or federated path is documented for it. The Sandboxes API is the exception: it takes a Nebius IAM token, which a workload could obtain keylessly through the federated credentials graded A on nebius-ai-cloud, but a side endpoint does not lift the product. The Nebius IAM StaticKey enum has an AI_STUDIO service, which hints that Token Factory keys are IAM static keys (six-month default, three-year cap); that is not stated in any Token Factory doc and is not credited. If Token Factory documented IAM-token auth for inference, this card would inherit the AI Cloud grade.

Outbound F

How this product reaches into your cloud, repos, or other SaaS.

Fine-tuning jobs export to Hugging Face with a pasted HF access token and to Weights & Biases with a pasted W&B API key; Sandboxes import images from private registries with a username and personal access token.

docs.tokenfactory.nebius.com checked 2026-10-22
“Hugging Face access token (PAT) with write access to `output_repo_name`.”
docs.tokenfactory.nebius.com checked 2026-10-22
“title: Api Key”
docs.tokenfactory.nebius.com checked 2026-10-22
“Authenticate with a container registry via Personal Access Token.”
docs.tokenfactory.nebius.com checked 2026-10-22
“**Working with custom model weights** is currently in beta and available on request.”

Checked: fine-tuning integrations (wandb, mlflow, hf), Sandboxes image import from Docker Hub, GHCR and other registries, custom model weights (beta, support-enabled, no public auth details) and dataset upload (pushed in by you, inbound). Every reach-in is a static token you paste. The Hugging Face guidance asks for write access to the one output repo, which on its own would be D; the W&B API key and registry PATs come with no least-privilege guidance, so the surface is F. No cloud bucket integration is documented.

Hygiene for any static credential the product still issues

✗ Expiry enforceable	The documented create flow asks only for a name. No expiry, default lifetime or org policy is documented for Token Factory API keys; the API keys page itself is behind a login. docs.tokenfactory.nebius.com checked 2026-10-22 “Enter the key name and then click Create.”
✗ Rotation via API	Keys are created and deleted in the Token Factory UI; the API reference (models, completions, files, fine-tuning, batch, datasets, endpoints, Sandboxes) has no key endpoints. docs.tokenfactory.nebius.com checked 2026-10-22 “The value cannot be opened again later.”
✗ Scoped keys	A key lives in one project, but project Members have full access to every project API, so a key is the creating user's project access. The groups page mentions service accounts in passing; no way to create one or give it a narrower role is documented in Token Factory. docs.tokenfactory.nebius.com checked 2026-10-22 “\| API Keys \| Create, Delete \| Create, Delete own API keys \|” docs.tokenfactory.nebius.com checked 2026-10-22 “A project is a logical container for your Nebius Token Factory resources.”
✗ Last-used visible	No last-used timestamp is documented; the API keys page is behind a login. docs.tokenfactory.nebius.com checked 2026-10-22
✗ Leak revocation	Nebius does not appear in GitHub's secret-scanning patterns table. The docs say compromised keys can be revoked automatically but name no scanning programme, so it fails per the Partner-column precedent. docs.github.com checked 2026-10-22 docs.tokenfactory.nebius.com checked 2026-10-22 “If a key is compromised, Nebius Token Factory can automatically revoke it”

Notes

Split from nebius-ai-cloud because Token Factory runs its own UI, projects and API keys rather than Nebius AI Cloud IAM; only its Sandboxes API shares the IAM token. Zero of five hygiene checks pass on public docs. The API keys page is login-walled, but outbound is F regardless, so the headline would not move. Accepting Nebius IAM tokens (and so federated credentials) on the inference and fine-tuning API would lift inbound to the AI Cloud grade.

Badge

[![wif.tax grade](https://wif.tax/badge/nebius-ai-studio.svg)](https://wif.tax/p/nebius-ai-studio/)