Documentation

Overview

Every pump.fun coin earns a creator fee on each trade. On most launchpads that fee goes to whoever launched the coin. Here it goes to compute for the coin's paired model instead: it is swept into a compute vault that belongs to that coin alone.

We say SI (super intelligence) rather than AI. An SI model is any model a coin can be paired with.

Quickstart

Launching takes about a minute and one wallet signature.

1. 1
   Connect a Solana wallet

   Phantom, Solflare, Backpack or any Wallet Standard wallet, or WalletConnect on mobile. You need enough SOL for network fees plus any initial buy.
2. 2
   Pick an SI model

   Choose a frontier model from the tiles, or search the full catalog. The pairing cannot be changed later.
3. 3
   Describe the coin

   Name, ticker and image are required; description and links are optional. The description is also given to the model as context.
4. 4
   Sign once

   One transaction creates the coin on pump.fun, with an optional initial buy in the same transaction. The coin page opens as soon as it confirms.

Models and pairing

The catalog is read live from OpenRouter and refreshed hourly, so new models become pairable the day they ship. A coin pairs with one concrete model, so the catalog leaves out batch-only variants, rolling aliases that point at whatever is newest, free tiers and routers.

Frontier flagships today

The newest top-tier model from each frontier lab. These are the tiles on the launch form and the models orbiting the home page.

Choosing a model

• Cheaper models go further. The same fees buy far more tokens on a fast, efficient model than on a flagship. A busy terminal on a frontier model can use up its compute quickly.
• Pairing is permanent. It is written into the coin's on-chain metadata. To pair with a different model, launch another coin.
• Many coins, one model. Any number of coins can pair with the same model. Each keeps its own compute; they do not pool.

Launching a coin

A coin launched here is an ordinary pump.fun coin: same bonding curve, same migration, tradable on pump.fun and anywhere else. One thing differs. Its creator is not your wallet but a compute vault created for that coin alone.

Fields

What the transaction does

The server pins your image and the coin's metadata to IPFS, generates the mint and its compute vault, and builds one transaction for your wallet to sign: pump.fun's create_v2 with the vault as creator, followed by a buy if you set one. Nothing is sent until you sign, and the coin is listed only once the transaction confirms.

Metadata

Alongside the standard fields, the metadata carries an si block, so anyone can read which model a coin is paired with and where its fees go:

{
  "name": "My Token",
  "symbol": "TOKEN",
  "image": "https://…/ipfs/…",
  "description": "…",
  "si": {
    "model": "anthropic/claude-fable-5.1",
    "computeVault": "<vault address>"
  }
}

Fees and compute vaults

pump.fun pays its creator fee to whichever account is the coin's creator. Because each coin's creator is its own vault, fees never mix between coins and never reach the launcher.

• One vault per coin. Generated at launch and shown on the coin page. Anyone can inspect it on Solscan.
• Swept automatically. Accrued fees are claimed into the vault continuously once at least 0.002 SOL is waiting. Each claim is a top-up, recorded with its transaction on the coin page and in the compute feed.
• Before and after migration. On the bonding curve fees arrive as SOL; after migration the AMM pays them as wrapped SOL. Both count toward compute.
• Accruing vs claimed. The coin page shows both: fees pump.fun owes the coin that have not been swept yet, and fees already in the vault.

How compute is counted

A coin's compute is every fee claimed into its vault, less what has been spent on prompts to its model.

Fees are paid in SOL while models are priced in dollars per token, so the two meet at the current SOL price. The token estimate assumes chat-shaped traffic, where a prompt and its context are about three times the length of the reply.

Worked example

A coin paired with Claude Fable 5.1 has had 2 SOL of fees claimed, and its terminal has spent $12.50. SOL is $119.95.

The SI terminal

Every coin page has a terminal: a public chat with the coin's paired model. Replies stream in as they are generated. The model is given the coin's name and description as context.

Metering

Each reply is metered at the model's OpenRouter price for the tokens it actually read and wrote, and that cost is drawn from the coin's compute. Before answering, the terminal checks the coin can cover a full-length reply; if it cannot, the terminal is disabled until trading adds more compute.

Limits

Coin lifecycle

1. 1
   Launch

   The coin is created on pump.fun with its compute vault as creator. Compute starts at zero.
2. 2
   Bonding curve

   The coin trades on pump.fun's curve. Every trade accrues a creator fee, and compute grows with volume.
3. 3
   Terminal opens

   Once compute can cover a reply, anyone can prompt the paired model.
4. 4
   Migration

   When the curve completes, the coin moves to pump.fun's AMM. Fees keep flowing into the same vault.
5. 5
   Ongoing

   Compute rises with trading and falls with use. Its whole history stays on the coin page.

Reading a coin

What each number on a coin card and coin page means.

The watermark behind a coin card is the logo of the lab behind its paired model. The lab filter above the grid narrows the list to coins paired with one lab's models.

API

Read-only JSON endpoints, no key required. Responses are cached for a few seconds.

{
  "coins": [
    {
      "mint": "…", "symbol": "TOKEN", "name": "My Token",
      "model_id": "anthropic/claude-fable-5.1", "model_name": "Claude Fable 5.1",
      "vault": "…", "total_fees_sol": 2, "compute_used_usd": 12.5,
      "marketCapUsd": 41200, "progress": 0.37, "graduated": false,
      "compute": { "sol": 1.8958, "usd": 227.40, "tokens": 11370048 }
    }
  ]
}

{
  "models": [
    { "id": "anthropic/claude-fable-5.1", "name": "Claude Fable 5.1", "lab": "anthropic",
      "context": 1000000, "prompt": 0.00001, "completion": 0.00005 }
  ],
  "flagships": ["anthropic/claude-fable-5.1", "openai/gpt-6-astra", "google/gemini-3.8-flash", …]
}

# request
curl -N https://si16z.fun/api/chat \
  -H "content-type: application/json" \
  -d '{ "mint": "<coin mint>", "prompt": "Summarize this coin in one line." }'

# errors come back as JSON with a status code
402  { "error": "$TOKEN is out of compute. Trading refills it." }
429  { "error": "too many prompts, try again in a minute" }

FAQ

Glossary