NockAI

Nock. Aim. Ship.

Nock Coder is an open coding model in 1 GB. It runs on the laptop you already have, on Windows, macOS or Linux, with no GPU, no account and no API key.

Early preview · Built on Qwen2.5 Coder · Apache 2.0

$NOCKAI CA Soon

The code first.

Ask for a snippet, a fix or a contract. Nock Coder answers with the code and at most one sentence. It is a 1.5B model and does not always get it right, so the results are below.

Example output

nock-coder-1.5b Local · GGUF

Everyday questions, answered fast

Up and running in a few minutes

Runs on Windows, macOS and Linux with Ollama. No account and no API key.

01

Get Ollama

winget install Ollama.Ollama

02

Download Nock Coder

About one gigabyte, downloaded once.

ollama pull hf.co/Nock-AI/nock-coder-1.5b-GGUF

03

Ask it something

Leave the quotes off to start a chat.

ollama run hf.co/Nock-AI/nock-coder-1.5b-GGUF "Reverse a string in Python"

Prefer an API?

Nock Coder is live on Surplus at $0.03 per 1M tokens, pinned to the exact commit it launched with. Use this model id with your Surplus key.

launch/Nock-AI/nock-coder-1.5b@18859c0

Will it run on my machine?

Pick your setup. This is a memory estimate for the 4 bit build, and we will add measured speeds after release.

  • Nock Coder, about 2 GB
  • Typical background apps
  • Free

Each square is half a gigabyte. Background use is an illustration, of your memory goes to the model.

Short answers. Your own machine. No account to create and nothing to send anywhere.

One model for each job

Each model starts from the base that fits its task and gets its own name. We ship them one at a time.

  • Nock Coder

    Writes and fixes code in many languages, including Solidity for Robinhood Chain.

    nock-coder-1.5bPreview

  • Nock Translate

    Translates crypto docs and wallet prompts into 8 languages and keeps the terms consistent.

    nock-translatePlanned

  • Nock Explain

    Describes what a transaction or contract does in plain language.

    nock-explainPlanned

What is inside Nock Coder

The plain facts, the same ones that go on the model card.

Base model
Qwen2.5 Coder 1.5B Instruct
Method
LoRA fine tune, two passes, about 1 percent of the weights adjusted
Training data
4,572 short, code first answers from three openly licensed datasets (MIT and Apache 2.0): general code, one line requests and Solidity
Format
4 bit GGUF, roughly 1 GB
Runs with
Ollama and llama.cpp on Windows, macOS and Linux
License
Apache 2.0

Questions & Answers

Short answers about the model, running it and licensing.

NockAI

01What is NockAI?

NockAI is a set of open language models for people who build on and use Robinhood Chain. The first one is Nock Coder, a coding model. More follow once it is tested.

02Which base model is it built on?

Nock Coder is a LoRA fine tune of Qwen2.5 Coder 1.5B Instruct, which is licensed under Apache 2.0. We trained it on short, code first answers.

03Is NockAI affiliated with Robinhood?

No. NockAI is an independent project. Robinhood and Robinhood Chain are trademarks of their owners and are named only to describe what the models are for.

Model & use

04Does it run on my computer?

Yes. We publish a quantized GGUF build for llama.cpp and Ollama. At 1.5 billion parameters it runs on a typical laptop, including one without a dedicated graphics card.

05How good is it?

It is a 1.5B model, so expect mistakes. In our test of 50 prompts, all 50 answers from v2.1 started with code, at about a quarter of the base model's length. On HumanEval it passes 118 of 164 problems, 6 more than the base model. On 15 Solidity tasks, 10 answers compile, against 8 for the base model. Review any code before you run it. See the full results.

06Can I trust its smart contracts?

No model output should go on chain without review. Nock Coder is a drafting aid. Have a qualified person audit any contract before you deploy it.

07Is it free?

Running it on your own machine is free. We release the weights openly and state the license on each model card. A hosted API is also available on Surplus at $0.03 per 1M tokens.

08What data was it trained on?

Short code first answers drawn from two MIT licensed instruction datasets, one general and one for Solidity, plus a few identity examples. Both datasets contain text generated by other language models, and the model card lists them.

You write the prompt.
Nock Coder writes the code.

The weights, the GGUF build and the model card live on Hugging Face. Download it and try it on your own machine.