Kila AI
Less talk. More code.
The open coding model for Robinhood Chain builders. It runs offline on your own machine: no account, no API key, no cloud.
1.5B parameters · based on Qwen2.5 Coder · Apache 2.0
$ ollama run hf.co/KilaAI/kila-coder-1.5b-GGUF "Write a viem public client for Robinhood Chain mainnet."
```ts
import { defineChain } from "viem";
export const robinhoodChain = defineChain({
id: 4663,
name: "Robinhood Chain",
nativeCurrency: { name: "Ether", symbol: "ETH", decimals: 18 },
rpcUrls: { default: { http: ["https://rpc.mainnet.chain.robinhood.com"] } },
blockExplorers: { default: { name: "Blockscout", url: "https://robinhoodchain.blockscout.com" } },
});
import { createPublicClient, http } from "viem";
const client = createPublicClient({ chain: robinhoodChain, transport: http() });
console.log(await client.getBlockNumber());
```Code first, words last.
Ask for a network config, a contract, a deploy command or a Hardhat setup. Kila Coder answers with the code and at most one sentence. It is a 1.5B model and will get things wrong, so check the test results.
import { defineChain } from "viem";
export const robinhoodChainTestnet = defineChain({
id: 46630,
name: "Robinhood Chain Testnet",
nativeCurrency: { name: "Ether", symbol: "ETH", decimals: 18 },
rpcUrls: { default: { http: ["https://rpc.testnet.chain.robinhood.com"] } },
blockExplorers: { default: { name: "Blockscout", url: "https://explorer.testnet.chain.robinhood.com" } },
});// SPDX-License-Identifier: MIT
pragma solidity ^0.8.24;
import {ERC20} from "@openzeppelin/contracts/token/ERC20/ERC20.sol";
import {Ownable} from "@openzeppelin/contracts/access/Ownable.sol";
contract MintableToken is ERC20, Ownable {
constructor() ERC20("Mintable Token", "MTN") Ownable(msg.sender) {}
function mint(address to, uint256 amount) external onlyOwner {
_mint(to, amount);
}
}export RH_RPC_URL=https://rpc.testnet.chain.robinhood.com # chain ID 46630
forge create src/Counter.sol:Counter \
--rpc-url $RH_RPC_URL \
--private-key $PRIVATE_KEY \
--broadcastimport type { HardhatUserConfig } from "hardhat/config";
import "@nomicfoundation/hardhat-toolbox";
const config: HardhatUserConfig = {
solidity: "0.8.24",
networks: {
robinhood: {
url: process.env.RH_RPC_URL || "https://rpc.mainnet.chain.robinhood.com",
chainId: 4663,
accounts: process.env.PRIVATE_KEY ? [process.env.PRIVATE_KEY] : [],
},
},
};
export default config;Everyday questions,
one line each
- ›Viem config for Robinhood Chain testnet
- ›Reentrancy guard in Solidity
- ›Debounce a function in TypeScript
- ›Parse a CSV file in Go
- ›Hardhat network config for mainnet
- ›Fix an off-by-one error
- ›Write a unit test for this function
- ›Capped token supply with OpenZeppelin
- ›Mintable ERC-20 with OpenZeppelin
- ›Mint function with an owner check
- ›Sort objects by a key in Python
- ›Foundry script to deploy a contract
- ›Regex for an email address
- ›wagmi config for Robinhood Chain
- ›Dockerfile for a Node app
- ›Merkle airdrop with OpenZeppelin
- ›Verify a contract on Blockscout
- ›Binary search in Java
- ›Rust struct with serde
- ›Flatten a nested list
- ›web3.py check for Robinhood Chain
- ›Git command to undo the last commit
- ›Add type hints to a Python function
- ›SQL query for the top 3 customers
How it works
Kila Coder is a 4-bit GGUF file on Hugging Face. It runs with Ollama or llama.cpp on Windows, macOS and Linux.
-
Free and open source. Pick your system.
Download the app from ollama.com/downloadcurl -fsSL https://ollama.com/install.sh | shDownload the installer from ollama.com/download -
About 1 GB from Hugging Face, downloaded once.
ollama pull hf.co/KilaAI/kila-coder-1.5b-GGUF -
Leave out the question to start a chat.
ollama run hf.co/KilaAI/kila-coder-1.5b-GGUF "Hardhat network config for Robinhood Chain testnet"
Built for
Robinhood Chain work
Plus the everyday code around it. Small jobs you would otherwise search for, answered on your own machine.
Chain setup
Network config for Hardhat, Foundry, viem, ethers and wallets: chain IDs, RPC and explorer.
Solidity drafts
Contracts and interfaces to read, test and have audited before you deploy.
Deploy and verify
Foundry and Hardhat commands, testnet first, keys kept in environment variables.
Token contracts
OpenZeppelin ERC-20s: mintable, burnable, capped, with permit or a Merkle airdrop.
Fixing errors
Paste the error and the line. Get a fix to try, as code.
Everyday code
Python, JavaScript, shell and SQL for the scripts around your project.
Will it run on
my machine?
Pick your setup. Memory is an estimate for the 4-bit build. Speed is measured on one Apple M5 so far.
Should fit.
Runs on the CPU. Speed: not measured yet.
Each square is half a gigabyte. Background use is an illustration, not a measurement.
One focused model
per job
Each model starts from a base that fits its task. We release one at a time, and only after it has been tested.
-
Kila Coder
Short, code-first answers for everyday code and Solidity, plus Robinhood Chain network configs and deploy commands.
-
Kila Explain
Plain-language summaries of what a contract or a transaction does.
-
Kila Review
Points at risky patterns in Solidity code for a human reviewer to check.
What is inside
Kila Coder
The plain facts, the same ones that go on the model card.
| Base model | Qwen2.5 Coder 1.5B Instruct (Apache 2.0) |
|---|---|
| Method | LoRA fine-tune, about 1.2% of the weights trained, 2 passes over the data |
| Training data | 3,672 short code-first examples: OpenCodeInstruct by NVIDIA (CC BY 4.0, synthetic), MBPP by Google Research (CC BY 4.0), Solidity by Example (MIT), and about 210 Robinhood Chain and OpenZeppelin examples written for this project with AI help and checked by script |
| Format | 4-bit GGUF (Q4_K_M), 986 MB, plus the LoRA adapter (74 MB) |
| Runs with | Ollama and llama.cpp, on Windows, macOS and Linux (tested on macOS so far) |
| Speed | About 105 tokens per second on an Apple M5 with 16 GB, in Ollama. Other machines not measured yet. |
| Test results | 100 held-out prompts and HumanEval, before and after training. See every number, including the weak spots |
| License | Apache 2.0, with dataset credits on the model card |
Questions
you might have
Short answers about the model, running it and licensing.
Kila AI
01What is Kila AI?
An independent project building open coding models that run on your own computer. Kila Coder is the first one: code-first answers for Robinhood Chain builders, offline.
02Which base model is it built on?
Qwen2.5 Coder 1.5B Instruct from the Qwen team (Apache 2.0), fine-tuned with LoRA on 3,672 code-first examples.
03Is Kila AI affiliated with Robinhood?
No. Kila AI is not affiliated with, endorsed by or sponsored by Robinhood. Robinhood Chain is named only to describe what the model is meant to help with.
Model & use
04Does it run on my computer?
Yes. The 4-bit build is a 986 MB file that runs with Ollama or llama.cpp on Windows, macOS and Linux, without a GPU. On an Apple M5 it writes about 105 tokens per second.
05How good is it?
On 100 held-out prompts it starts every answer with code (the base model: 33 of 100) and passes 15 of 20 Robinhood Chain fact checks (the base model: 1). It scores 120 of 164 on HumanEval. See all the results.
06How should I use the contracts it writes?
As a first draft. Read and test every line, try it on testnet, and have a qualified person audit any contract before it goes to mainnet.
07Is it free?
Yes. The weights are released under Apache 2.0, which allows personal and commercial use. There is no account and no API key.
08Does it send my code anywhere?
No. The model runs on your machine through Ollama or llama.cpp. Kila AI runs no server that sees your prompts.
You ask the question.
Kila writes the first draft.
The 4-bit build, the LoRA adapter and the model card with every test result are on Hugging Face.
View on Hugging Facesoon