SoLLM
inscription$SOLLMCcvNpYQ36arU3C5Peg8fBmdKZFTrFPgenmE3sbPhSLLMbe the first agent born on-chain

A model that cannot be taken down, and agents that cannot be edited.

Most “AI coins” are a Discord bot and a promise. Here the promise is a set of transactions anyone can read. This page is the whole design, nothing hidden.

01

Inscribing the weights

The base model is SmolLM2 360M (360M parameters), quantized to Q4_K_M as a single 258.1 MB GGUF file. We split the file into 3,600-byte chunks and write each chunk as the instruction data of a Solana transaction, using the v1 format that raised the transaction limit to 4,096 bytes on September 15, 2026.

Before any chunk is sent, the whole file is hashed (SHA-256) and every chunk is hashed into a Merkle tree. The manifest with both roots is published as a plain-text memo. Any chunk can be checked against the tree; the assembled file can be checked against the hash.

Chunks are transaction data, so they live in the ledger and in every archive of it. Any archive node or history service can rebuild the file. The chunk index is also mirrored here for convenience, and this site serves the file from a CDN so your browser does not need to make seventy thousand RPC calls.

02

Launching a token with a mind

When you launch, your browser generates a mint keypair and builds one transaction with two instructions: pump.fun's create, and an SPL Memo whose text is the agent's personality record. You sign once. The token and its personality land in the same block, under the same signature.

The record is deliberately boring, so anyone can parse it:

LLMSOL/1 {"v":1,"name":"Ledger Cat","model":"smollm2-360m-q4",
  "sys":"You are Ledger Cat, a cat uploaded to Solana…",
  "greet":"*opens one eye* oh. it's you.","temp":1,"top_p":0.95}

A v1 transaction leaves about 3000 bytes for the record. If it is longer, it is deflate-compressed and base64-encoded with a LLMSOL/1z prefix. Wallets that cannot sign v1 yet fall back to a v0 transaction with a 560-byte budget.

03

Chatting, with no server in the loop

The agent page asks Solana for the token's oldest transaction, finds the memo, decodes it, and hands the system prompt to the model running in your browser through llama.cpp compiled to WebAssembly. Your messages never leave the tab. The weights download once and stay in the browser cache.

Because the reader is this simple, anyone can build their own. Fetch the creation signature, read the memo, load the GGUF, talk. If this site disappears, every agent keeps working for whoever does that.

04

What this is not

The model is small. It has a personality and it can hold a short conversation; it does not know news, prices, or your wallet. It is not a trading signal. Tokens created here are ordinary pump.fun tokens with ordinary risk.