# Compute API

Paid AI jobs on the network's Macs through one OpenAI-compatible endpoint, paid in USDC. A pilot, open to every MLX Wallet since 27 Oct 2026.

## What it is

Until the pilot, every job a Mac ran was verification work: the network chose the prompt, and nobody paid for the answer. The Compute API lets you send your own prompts. Your prompts run on Macs with dedicated seats, you pay per token in USDC, and the share of the network's tokens that served a paying request is public ([GET /api/compute/work](/network/compute)).

- **One endpoint**, OpenAI-compatible: `https://api.mlx.app/v1` (also served at `https://mlx.app/api/v1`). An OpenAI SDK works once you change its base URL.
- **One small open model** to start: `qwen2.5-0.5b-instruct` (mlx-community/Qwen2.5-0.5B-Instruct-4bit), the model every Mac already has loaded.
- **$0.02 per million tokens**, prompt and completion alike, from API credit you add from your MLX Wallet's USDC.
- **Checked by a second Mac.** Every job runs on 2 Macs of one GPU group and of different wallets. You get the answer they agree on, marked `mlx.verified: true`.
- **Paid to the Macs in USDC**, once a day, never locked: 90% of the price, split between the Macs that ran it. The other 10% goes to the treasury. A seat's MLX mining rate doesn't change.

:::cards
- [Quickstart](/compute/quickstart) A key, some credit and your first request.
- [Chat completions](/compute/chat-completions) What the pilot's endpoint takes and returns.
- [Credit and billing](/compute/credit-and-billing) API credit on the ledger, the daily bill, getting credit back.
- [How jobs run](/compute/how-jobs-run) Hand-out to two Macs, agreement and the Macs' pay.
:::

## What it isn't

> [!WARNING]
> **It isn't private.** Your prompts are sent to other people's Macs. Send only work you'd be happy for a stranger to read: evaluations, synthetic data, classification, rewriting public text. Never send personal data, secrets or anything confidential.

- **It isn't a chat product.** Each request waits while Macs pick the job up. That usually takes a few seconds, and up to 45 seconds before the request gives up. There's no streaming.
- **It isn't a big model yet.** One 0.5B model, greedy decoding, one user turn, up to 120 tokens out. Bigger models and longer answers need a newer MLX Node.

## Getting started

Any MLX Wallet can use the pilot. Sign in to the [console](/console) to make a key and add credit, then follow the [quickstart](/compute/quickstart). Questions, or work you'd like to run at scale: [contact@mlx.app](mailto:contact@mlx.app?subject=Compute%20API%20pilot).