Docs / Buying

Limits and errors

How many requests we take at once, how long we wait, and what each error means.

Concurrency

Our seller serves up to 30 requests at the same time. A 31st waits in the proxy until a slot frees. Long agentic sessions with large contexts hold a slot for the whole response.

Timeouts

We wait up to 10 seconds to connect to Venice for a request, and retry once on a connection error. A model that takes long to think is not a timeout: once Venice has accepted the request we wait for the answer.

Context and output

The context limit is the model's; a request over it gets Venice's 400 back. For some models Venice charges output tokens past the request's max_tokens at a higher rate, or has no hard ceiling on output; we pass those prices through. Set max_tokens when the cost of a runaway answer matters to you.

Errors you may see

Status Text contains What it means What to do
400 unknown parameter, Reasoning is mandatory Venice refused a field, or the model needs a setting the request did not send Read the body; drop the parameter or set what it asks
402 insufficient Your buyer balance or channel cannot cover the request Fund the buyer, or close idle channels to free reserve
404 model not found The name is not in our catalog today Check the table on Use for the current names
403 free_budget_exhausted The free tier's daily total is used up Retry after 00:00 UTC (the retry-after header says when), or use a paid model
429 Venice rate-limited us for that model Wait and retry; the proxy may pick another seller if you are not pinned
502 Pinned peer ... is not reachable Your proxy is pinned to us and cannot find us right now Wait a minute and retry; a fresh proxy may still be discovering peers. If it persists, update the AntSeed CLI and restart the proxy. Otherwise we may be down
502 does not offer provider or does not support this request Pinned to us, but we do not serve that model or that request shape Pick a model from the catalog, or clear the pin
502 pinned peer could not complete the request We answered with an error (Venice failed, or the key ran dry at that moment) Retry; if it repeats within minutes, our status is probably showing it in the live feed

An error from Venice is passed through unchanged, with its status and body.

Free models

Five models are served free: gpt-oss-120b, mistral-small-3-2-24b-instruct, google-gemma-4-26b-a4b-it, google-gemma-4-31b-it and nvidia-nemotron-3-nano-30b-a3b. Text only, with tools; the free tier does not take images. The allowance is shared by the five: 200 requests per wallet and 400 per IP address in a rolling 24 hours, with up to 16,000 tokens of context per request and 1,024 tokens of output (2,048 on gpt-oss-120b). Above the size the request is refused with a 400 that says so and names a paid model that takes longer prompts; above the count, with a 429 that names the limit and when it resets. The free tier also has a daily total across all buyers, and one wallet can use up to a tenth of it. Once the total is used up, free requests get a 403 free_budget_exhausted until 00:00 UTC; once a wallet has used its share, that wallet gets a 403 free_wallet_budget_exhausted until 00:00 UTC. Both come with a retry-after header. Paid models are not affected. Free requests are metered like paid ones, but there is no receipt with a cost and they do not appear in the live feed.

More free requests with ANTS

A wallet that locks ANTS in our seller pool gets a larger daily allowance, counted on its stake rather than on the wallet or the IP address. The AntSeed client reads the tiers from our listing, and a 429 from us names them too.

ANTS locked for a year Staking power Free requests a day
none 200
about 5,000 260,000 1,000
about 25,000 1,300,000 2,000
about 100,000 5,200,000 5,000
about 250,000 13,000,000 10,000

Staking power is ANTS multiplied by the weeks left on the lock, so a shorter lock needs more ANTS for the same tier, and a position counts from the epoch after it is made. The stake also raises the pool's weight, which multiplies the points your paid requests earn. The tiers are the same on every free model, and there is no unlimited tier. Every wallet, staked or not, also stays within its share of the free tier's daily total. Staking ANTS is done in the AntSeed client or app, under the seller's pool; it is AntSeed's contract, not ours.

Prices climbing

Late in a busy day our price for a model can exceed the price limit you set in your proxy. If you are not pinned, the proxy routes elsewhere; if you are, the request fails with a 402-style message from the proxy. See Prices through the day.

When we are down

Software fails, and ours does too. When it does, the live feed goes stale within three minutes, the numbers on the home page stop moving, and pinned requests get the 502 above. We do not have a status page yet; the feed is the closest thing.

Last updated 2026-11-13 · this page as Markdown · llms.txt