Limits

Every limit on this page is generated from the same definitions the service enforces. There is no second, friendlier set of numbers.

Quotas

Every approved account starts on the academic tier. Sustained use above 3× the tier's median raises an alert rather than a block.

Counter	Window	academic tier
tiles	day	50,000
rows	day	200,000
jobs	day	5
jobs_month	month	30

A window is a rolling window, not a calendar one. When a counter is exhausted the answer is quota_exceeded, and it says which counter and when it resets.

What charges what

Input kind	Counter	Charged
patches	tiles	one per item
embeddings	rows	one per item
bulk_rna	—	nothing
tissue	—	nothing
disease	—	nothing

Work is charged by what causes it. Tiles charge per tile and embedding rows charge per row, so a small archive of large tiles and a large archive of small tiles cost what they cost to run rather than what they weigh. Covariates charge nothing.

Sizes

Input kind	Item cap	Upload cap	Per-model bounds from
patches	50,000	2 GiB	card.input_spec
embeddings	50,000	256 MiB	card.input_spec
bulk_rna	—	64 MiB	card.output

The item cap is the point above which a submission is refused rather than subsampled — see Observations. The upload cap is enforced before the bytes are written, so an oversized upload fails fast with payload_too_large.

Per-model bounds are narrower than these, and they come from the model's own input_spec. These are the outer bounds of the service; a model may need less.

Request rate

Bucket	Limit
read	600 per minute
write	60 per minute
submit	10 per minute

Rate limits are per key, not per address, so a lab behind one address is not one client. Every response carries the state of your buckets:

RateLimit-Limit, RateLimit-Remaining, RateLimit-Reset, Retry-After

Read those headers instead of guessing. When you do run out, the answer is rate_limited with an interval to wait, and waiting the interval is the correct response — retrying immediately makes it worse for everyone including you.

One job at a time

One job per key is in flight. A second submission is queued, and its position and estimate are in the job response. This is a property of a small amount of hardware being shared fairly, not a tier you can buy out of.

If a limit is wrong for you

Limits are set from what the service can actually deliver, and they are tuned from real usage rather than guessed once. If your work does not fit, say what you are trying to do — the answer may be a raised counter, and it may be that the local path is a better fit than the hosted one.