A market for tokens, not a rental desk

Your computer can earn.
Your apps can stop overpaying.

People with a spare GPU serve open models. People who need AI pay by the token. Scalar matches them, keeps score, and sends the money. You never open a port.

Get an API key Start earning

One command to provide. One key to buy. Swap base_url and keep the OpenAI SDK.

How you know it is the real model

Signed catalog

Every official weight file is hashed with SHA-256 and signed with Scalar’s Ed25519 key. Providers may only advertise what is in that catalog.

Check on join

Before a machine is listed, the agent hashes the file on disk (and a few random chunks) and we compare it to the catalog. A mismatch never goes live as verified.

Random polls

We keep asking providers to re-prove the hash plus a fresh nonce. Fail twice and they drop off the market. Every reply carries a receipt you can check at /v1/receipts/<job>.

Buy inference

Create a key. We start you with a little credit.
Point any OpenAI client at http://127.0.0.1:8787/v1.
We route to the cheapest healthy machine that has the model.

from openai import OpenAI
client = OpenAI(
    base_url="http://127.0.0.1:8787/v1",
    api_key="sk-scalar-YOUR-KEY",
)
print(client.chat.completions.create(
    model="Qwen3.8-27B",
    messages=[{"role":"user","content":"Hello from Scalar"}],
).choices[0].message.content)

Earn from a PC

Install nothing fancy. If you already run llama.cpp, Ollama, or vLLM, we use it.
Run one command. The agent phones home. No port forwarding.
You keep 85% of every token you serve.

curl -fsSL http://127.0.0.1:8787/install.sh | bash

Live market

See all prices

Qwen3.8-27B

$0.20 / million input tokens

1 online · routable

Create your key