Is Replicate Down?

No — Replicate is up

Reachable from all 8 checked regions

Average response time: 786ms

Last checked 3 hours ago · checks run every 6 hours

Official status page: https://www.replicatestatus.com

Replicate uptime

99.57%

Last 7 days

99.67%

Last 30 days

99.67%

Last 90 days

770ms

Avg response, 30 days

Measured from multiple regions every 6 hours. Percentages count only checks that returned an availability answer — 10 days measured so far. A dash means that window does not yet hold enough measured days to publish a figure.

30-day history

2-day clean streak

Jul 25 Today

No downtime Partial Downtime Not measurable No data

Reachability by region

Each region runs its own request from a different part of the world. A service can be up for one continent and down for another, which is usually the first sign of a routing or CDN problem.

ams

953ms

DNS 0ms TCP 2ms TLS 9ms TTFB 939ms

arn

763ms

DNS 0ms TCP 1ms TLS 12ms TTFB 750ms

bom

1201ms

DNS 0ms TCP 3ms TLS 12ms TTFB 1183ms

nrt

672ms

DNS 0ms TCP 1ms TLS 7ms TTFB 664ms

ord

368ms

DNS 0ms TCP 3ms TLS 8ms TTFB 354ms

sin

714ms

DNS 0ms TCP 1ms TLS 7ms TTFB 706ms

sjc

676ms

DNS 1ms TCP 3ms TLS 9ms TTFB 663ms

syd

1317ms

DNS 0ms TCP 1ms TLS 6ms TTFB 1308ms

yyz

303ms

DNS 0ms TCP 1ms TLS 7ms TTFB 295ms

What Replicate does

Replicate runs machine learning models behind an API, so applications can call image, audio and language models without operating GPUs themselves. Models are packaged as containers and pushed to its registry at r8.im. Products that generate images or transcribe audio often depend on it for a user-facing feature rather than a background job, which makes an outage immediately visible.

What an outage looks like

Calls to create a prediction return errors, or are accepted and then never complete, leaving predictions queued indefinitely. Cold starts stretch well past normal, so a model that usually answers in seconds times out instead. Pushes to the r8.im registry fail during a deploy. The dashboard can load while inference is failing, which hides the problem from a quick check.

What to do about it

Check replicatestatus.com, which separates the API, inference and training, the r8.im registry and the website, so a failing deploy and a failing prediction are told apart. Predictions are asynchronous, so poll rather than assuming a stalled call has failed. Applications calling Replicate on a user-facing path should degrade or fall back rather than hold the request open.

Is it down for everyone, or just you?

If this page says Replicate is up but it is not loading for you, the problem is between you and them. Run a check against any URL from all 18 regions to find out where it breaks.

Test it yourself

Related services

Replicate outage FAQ

Do I get charged for predictions that fail?

Replicate bills for compute time used, so a prediction that never ran should not accrue meaningful cost, while one that ran and then failed to return may. Billing questions arising from an incident are worth raising with support rather than assumed either way, and support is itself a component that can be affected during a large outage.

Why is the first call after an idle period so slow?

Models that are not kept warm have to start a container and load weights onto a GPU before they can answer, so the first request after an idle period is far slower than the ones after it. During an incident that delay stretches further, which is why a timeout is often the first symptom of trouble rather than an explicit error.

Does an outage affect deploys as well as inference?

They are tracked separately. The r8.im registry has its own component, so pushing a new model version can fail while existing predictions run normally, and the reverse happens too. A deploy that fails during an incident is usually safe to retry once the registry recovers, since a partial push does not replace the running version.

Should my app call Replicate synchronously?

Predictions can take anywhere from a second to several minutes depending on the model, so holding a user request open for one is fragile even on a good day. Using webhooks or polling, with a visible pending state in the interface, means an incident degrades that feature rather than breaking the page built around it.

How we measure this

We request Replicate's public endpoint every 6 hours from Fly.io regions across six continents — 8 of them answered the most recent check.
A region counts as down only when it gets no usable HTTP response. A 403 or 429 means the origin answered and refused us, which we report as blocked, never as an outage.
A single failing region is treated as probe noise. We only change the verdict when two consecutive cycles agree.
Response times average only the regions that actually served the page, so a timeout never inflates the number.
Where Replicate publishes an official status feed we read it too. An all-clear from the vendor can soften an unconfirmed degradation; a vendor-declared outage only worsens our verdict when our own checks corroborate it.

Get alerted when Replicate goes down

This page refreshes every 6 hours. Your own monitors run as often as every 30 seconds, from the same 18 regions, and tell you the moment something breaks.

Start Free Monitoring

Free plan available No credit card required