Limits

Your organisation’s tier sets most of its limits. A new organisation starts at tier 0, and buying credit moves it up. Every key and browser session of the organisation shares its limits.

Tiers

These are the standard tiers, and the credit you buy to reach each:

Tier (credit bought)	Requests a minute	Input tokens a minute	Operations at once
0 ($0)	30	100,000	1
1 ($10)	120	1,000,000	4
2 ($100)	240	2,000,000	8
3 ($1,000)	480	4,000,000	16
4 ($10,000)	3,000	8,000,000	32

Each tier also sets how many of your fine-tuned models can answer calls and stay warm at once, and the largest dataset form you can upload:

Tier	Your models answering	Your models kept warm	Dataset form
0	1	0	16 MiB
1	3	0	32 MiB
2	5	1	32 MiB
3	10	3	64 MiB
4	20	5	64 MiB

Score many texts shows what these limits mean for a long run.

Move up a tier

Your tier comes from all the credit your organisation has bought, before tax. Spending credit, or letting it expire, never lowers it. Starter and promotional credit don’t count towards it, and a refund or a dispute takes its amount off the total. Each organisation’s tier comes from its own purchases.

Buy credit in Billing in the console (buy credit).
The new tier applies once the payment is verified, from your next request.

next_tier on GET /v1/account says how much more credit reaches the next tier (tier and limits). To evaluate OneHot at a higher tier before you buy, for a one-off large run, or for limits above the top tier, write to help@sqwish.ai.

Account limits

Read your tier and its limits on GET /v1/account:

from sqwishai import OneHot

with OneHot() as client:
    account = client.account.get()

print(account["tier"], account["effective_limits"])

import { OneHot } from "sqwishai";

const client = new OneHot();
const account = await client.account.get();

console.log(account.tier, account.effective_limits);

curl --fail-with-body https://api.sqwish.ai/v1/account \
  -H "Authorization: Bearer $SQWISH_API_KEY"

tier is the tier your organisation is on, and effective_limits holds its limits. An inference key can’t read this route (Access levels).

Field in `effective_limits`	Limit
`rpm`, `tpm`	Requests and input tokens a minute. An image counts the tokens the model reads for it, in `usage.image_tokens`. These are the same tokens it bills
`concurrency`	Operations running at once
`read_rpm`	Reads a minute, including replays of existing operations
`write_rpm`	Writes a minute, other than decisions and their estimates
`estimate_tokens_per_minute`	Tokens a minute for estimates
`stored_objects`, `stored_mib`	Datasets, labelling sources, fine-tunes and labelling jobs the account keeps, and the space they take (Places and storage)
`upload_bytes`	Bytes in one dataset form, its JSONL files together. Higher tiers take more (Request size)
`answering`	Models that answer calls at once (Switch a family or a model on)
`kept_warm`	Models kept warm at once (Keep a model warm)
`fine_tunes`	Fine-tunes queued, running or cancelling at once (Fine-tune a model)

The tiers tables give the limits that grow with your tier. The others are the same at every tier. The console’s Account page explains them all. These are the standard values of the limits that are the same at every tier:

Limit	Value	Field in `effective_limits`
Reads a minute	120	`read_rpm`
Writes a minute, other than decisions and their estimates	60	`write_rpm`
Input tokens a minute in estimates	1,000,000	`estimate_tokens_per_minute`
Datasets, labelling sources, labelling jobs and fine-tunes you keep	200	`stored_objects`
The space they take, in MiB	2,048	`stored_mib`
Fine-tunes queued, running or cancelling at once	4	`fine_tunes`

Rate limits

/v1/account reports your tier’s limits on requests and input tokens a minute, and on operations running at once. An image counts the tokens the model reads for it. Reads and estimate tokens have separate budgets. Authenticated reads use the read budget. So do replay attempts of existing operations, pending or conflicting ones included. Per-address limits also apply across dynamic routes, before expensive processing.

Over a limit, a request gets 429 with Retry-After. Its message names the limit and when it resets. It names the limit as effective_limits does: rpm, tpm or concurrency.
A request with more input tokens than your account allows in a minute gets 422 too_many_tokens. No wait would admit it.
When the rate limiter’s storage is unavailable, a request gets 503.

Revoking a key, or turning automatic top-ups off in the console, never waits for these budgets.

Input tokens

Every model, fine-tunes included, takes up to 65,536 input tokens a request. Tokens are counted as they are billed, in usage.input_tokens, images included. /v1/models gives this limit as each model’s max_input_tokens. A longer request is refused with 422 context_too_long before any model reads it. Nothing is charged. Rarely, a shorter request gets the same error too. Split it.

Request size

Body	Limit
JSON body	Up to 1 MiB. A decision with images can be up to 48 MiB: only its images may take it past 1 MiB (images)
Dataset JSONL files, in a multipart form	Up to your tier’s `upload_bytes`
Labelling source	Up to 16 MiB and 2,000 rows

A body over its limit gets 413. Send less. A dataset form over upload_bytes gets 413 request_too_large, and its message names the tier that takes it. You can also buy credit to reach that tier. A compressed body must fit these limits both as sent and decompressed (Compression).

Rows and decisions

What	Limit
Rows in a dataset, across its versions	1 to 100,000
Training rows in the dataset of a fine-tune that trains every row: a family’s first, or one with no model yet	100 to 100,000
Training rows for a prompt tune, asking its decision	50 to 1,000
Decisions in a native request	1 to 128 in the schema. The model you choose may allow fewer. Each model’s `max_questions` in `/v1/models` is the most it allows
Decisions in the playground	At most 8

The row ranges are the limits in /v1/onehot. Each *_rows value is the most, and each *_min_rows value is the fewest.

Places and storage

places says how much of answering and kept_warm the account uses. Each model counts once, however many switches make it answer or keep it warm. storage says how many datasets, labelling sources, fine-tunes and labelling jobs the account keeps, and the bytes they take. Archived ones count, and deleted ones don’t. A new one past either allowance is refused with 409 storage_allowance. A fine-tune stops counting once you delete its model, or once a rollback discards that model (Delete a model).

Get started

Make decisions

Improve accuracy

Use the API

Account and billing

Privacy and security

Integrations

Tiers

Move up a tier

Account limits

Rate limits

Input tokens

Request size

Rows and decisions

Places and storage

​Tiers

​Move up a tier

​Account limits

​Rate limits

​Input tokens

​Request size

​Rows and decisions

​Places and storage

Tiers

Move up a tier

Account limits

Rate limits

Input tokens

Request size

Rows and decisions

Places and storage