Who runs this, where it runs, and what happens to your request.
We publish the operating entity, the processing scope, who touches a request, and what is left of it afterwards.
There is a company behind the endpoint.
A contract with an API is a contract with whoever operates it. This is that legal entity, its jurisdiction, and the address a subpoena or a data-subject request would go to.
You can see where your own request ran.
The fleet changes, so this page does not publish an inventory of it. It publishes the part that is about you: which regions a request passes through, who operates them, and where your specific call ran.
rendered from the same register the DPA is built from · the global edge is transit and hosting, not model execution
The request path includes a qualified edge hop.
Ordinary self-hosted text inference keeps prompts and completions in memory. H3 encrypted prompts and optional MiniMax processing have separate retention rules in the Privacy Policy. Cloudflare may process decrypted content at TLS termination, depending on the deployed zone and edge configuration.
TLS from your machine to the edge.
Cloudflare may process decrypted content at TLS termination, depending on the deployed zone and edge configuration.
Key checked, limits applied, metadata recorded.
Ordinary self-hosted text stays in memory; H3 and optional MiniMax processing have disclosed exceptions.
Self-hosted ordinary text is returned and released; H3 ciphertext and media have separate retention rules.
Two categories, two different promises.
aligned with the Privacy Policy and DPAOrdinary self-hosted text inference is processed in memory without persistence. H3 prompts, rewrites and MiMo drafts are encrypted in PostgreSQL and erased within 7 days; encrypted backups rotate out within 30 days and are outside database erasure. Media results are kept for 7 days and uploads for 48 hours. Optional MiniMax processing and future 2K have separate disclosures in the Privacy Policy.
Metadata examples include token counts, timestamp, latency, model, endpoint/status/channel, and billing or abuse identifiers. The policy ceiling is up to 90 days.
Everyone who touches a request, and what they touch.
The tag on each row is the thing worth knowing: whether that party can see request content, or only metadata about it.
Locations, retention authority and terms are in the DPA. New or replacement parties are published for at least 7 days before processing, without separate notice. Decline optional MiniMax processing per order or contact us to disable it for your account.
What we can prove, and what we cannot yet.
We suggest checking these three things. Latency and throughput show 24-hour real-usage medians once there is data, and read as being measured until then.
Recorded per model from the day it went live, with planned maintenance windows marked separately from incidents and two availability figures rather than the flattering one.
Medians from real traffic over the last 24 hours. Until this model has data in that window, no number is shown here, and no controlled benchmark is claimed.
Public status observations are available. A complete persisted incident ledger is not yet claimed; planned maintenance is shown separately when supplied by the feed.
Every request is checked against who owns it.
An API key passes three checks in order and is refused at the first one it fails. Once through, it can only reach its own account's data.
This page is a plain-language summary. Every claim on it is meant to be a weaker or equal statement of what the legal documents say — if you find a place where this page promises more than they do, that is a defect and we want to hear about it.