Free inference

Use it now

This is a complete, working request. Nothing has been removed from it.

curl https://api.murakumo.cloud/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"murakumo-main",
       "messages":[{"role":"user","content":"hello"}]}'

OpenAI-compatible, so most existing clients work by pointing their base URL at https://api.murakumo.cloud/v1 and sending any string as the API key, or none at all. API documentation.

And an image, on the same terms

One image model is free too: awai-network/hokusai, the fleet's WAI checkpoint. Also complete, also nothing removed.

curl https://api.murakumo.cloud/v1/images/generations \
  -H 'content-type: application/json' \
  -d '{"model":"awai-network/hokusai",
       "prompt":"a lighthouse on a rocky coast at dawn",
       "size":"768x768"}'

Returns {"data":[{"b64_json":"…"}]}, one image per request. Measured 2026-09-10: about ten seconds for 768×768. Both models are listed by /v1/models , so an agent can discover them before it calls them.

Who this is for

People for whom the barrier to AI tooling was never the price of a token. It was the card, the phone number, the billing address, the bank that will not issue, the country the payment form rejects, or the twenty dollars a month that is not available every month. Those are the barriers this removes, and they are removed for everybody at once because there is nothing to check.

It is also for agents. A bot that has no funding can run here: an unauthenticated agent loop is an ordinary caller. The bot runtime descriptor carries the same terms in machine-readable form.

What is not promised

Speed. This runs on a small owned fleet, and it is slow compared to a commercial frontier endpoint. The current measured decode rate and the real concurrent slot count are published, per model, at /v1/models — measured, dated, and stated with the method used. They are not repeated here, because a number written into a page keeps being read after it stops being true.

Availability. When every slot is busy the service answers 429 and asks you to come back, rather than accepting the request and making you wait without saying so. A refusal you can see is better than a queue you cannot.

Confidentiality beyond the operator statement. Requests pass through the gateway and the selected backend. See the privacy policy before sending anything that needs a specific contractual control.

Donate

The fleet is donated by Jun Kawasaki. Donations do not go into a general fund: what money buys here is serving capacity, which is the one thing that decides how many people can be served at once and how fast. A donation buys slots.

This is sponsorship, not charity. murakumo.cloud is operated by AWAI Network, L.L.C., a Delaware LLC. It is not a tax-exempt organisation, there is no 501(c)(3) determination, and nothing you send is tax deductible anywhere. If you need a deductible donation, this is not one.

This is the operator's existing treasury, the same address paid inference settles to over x402. That is deliberate: one address with a public ledger can be checked, while a separate donations address would have to be reconciled against the first before anyone could say what a donation funded.

There is no fiat card path yet. If you want to sponsor capacity and cannot send USDC, write to the operator.

If you would rather pay

Paid access exists and is unchanged: x402 per-request payment settles 0.01 USDC on Base and needs no account relationship either. Paying does not buy priority over the free lane. It buys the dedicated heads, their larger context windows, and models this pool does not serve.

There is also murakumo Go — free inference in the browser, funded by one visible first-party ad rather than by the donation. Prompts are never shared with advertisers. Use whichever suits you; neither costs money.