01What is North Mini Code and How Does It Work?

North Mini Code is a free, local, and open-source coding model that you run on your own computer. It is trained specifically for writing software and works offline with no internet needed. The model is small but powerful: it has 30 billion parameters but uses a mixture-of-experts design.

This means only 8 of its 128 experts activate for each word it generates. Because of this, it runs very fast and yet scores higher on coding benchmarks than models four times its size.

In this episode, Julian Goldie showed how North Mini Code works inside an agent operating system. You type a request, like "Build a colorful to-do list app with voice control," and the model writes the code in seconds. You can then preview the app live.

Everything stays on your machine, so your data is private. The model is also available on Ollama, making it easy to pull and use.

02Key Features of North Mini Code

North Mini Code comes with several standout features that make it useful for real-world coding:

  • Mixture of Experts: Only 8 out of 128 experts fire per word, so it runs fast on a laptop. Julian measured 92 words per second on an M4 Max Mac.
  • 256k Context Window: The model can handle long files, similar to other top models like Kimi K2.7.
  • RLVR Training: It was trained with reinforcement learning and verifiable rewards on real software and terminal tasks, not just benchmark puzzles.
  • Native Tool Use: It is built to drive coding agents, call tools, run a terminal, and think between steps.
  • Apache 2.0 License: The model is free and open source, so you can use it for commercial projects too.
  • Voice Control: You can build apps by speaking commands, as Julian demonstrated with the to-do list app.
  • Offline and Private: Nothing ever leaves your machine, and you can work without an internet connection.

03How to Set Up North Mini Code

Getting North Mini Code running takes about 10 minutes. You need Ollama installed on your computer. Follow these steps shown by Julian in the episode:

  1. Update Ollama: Make sure you have the latest version of Ollama installed.
  2. Pull North Mini Code: Run the terminal command ollama pull north-mini-code (or the exact name provided).
  3. Keep the Model Warm: Once loaded, keep it in memory so it responds instantly.

    Julian does this by running it on a loop inside the terminal.

  4. Wake It Once: After the initial load, the model stays ready for repeated use with no lag.
If you want the full coding engine with voice input, live previews, and saved workspaces, Julian offers it inside the AgentOS dashboard at airpufferboardroom.com. But you can also run North Mini Code alone using just these three terminal commands.

04Using North Mini Code to Build Apps

Julian walked through a live example to show how North Mini Code works. He opened the local section of his agent operating system and selected North Mini Code as the model. Then he typed a prompt: "Build a colorful to-do list app and control this with our voice." The model started coding immediately.

After a few seconds, the HTML was ready and he could preview the app directly inside the workspace.

  • The app worked correctly and responded to voice commands.
  • Every app built with the model is saved in the workspace, so you can go back to it later.
  • You can also plug North Mini Code into other agents like Hermes agent. Julian showed it running inside Hermes as a local agent.
  • You can toggle thinking mode on or off. With thinking off, a simple snake game built in about 20 seconds.

    With thinking on, it took about 48 seconds but produced deeper reasoning.

05Benchmark Results and Performance Comparisons

In the episode, Julian showed benchmark data from the artificial analysis coding index. North Mini Code scored 33.4, which beats several larger models:

  • Outperforms Gemma 4 on Terminal Bench, Terminal Bench Hard, and SW Bench Verified.
  • Outscores Nemetron 3 Super, Mistral 4, and DevStrawl 2.
  • Runs at 92 words per second on Julian's M4 Max Mac.
  • The model file is about 19 gigabytes at Q4 quantization, so it fits on a 16‑ to 36‑gigabyte Mac with room to spare.
Compared to Qwen 3.6, North Mini Code is not quite as strong, but Qwen 3.6 requires a more powerful setup. For a model that runs locally on a laptop, North Mini Code punches far above its weight. Julian also noted that you should test the model yourself rather than trusting benchmarks completely.

06Tips, Limits and Common Mistakes

Julian shared several practical tips and warnings for getting the best out of North Mini Code:

  • Keep the model warm: If you let it unload from memory, you will face lag. Keeping it warm makes it feel instant.
  • Choose speed or depth: Turn thinking mode on when you need deeper reasoning (e.g., complex logic) and off when you want fast results.
  • File size matters: The Q4 quantized version is 19 GB. Make sure you have enough free RAM (16–36 GB recommended).
  • Don't believe benchmarks blindly: Julian encourages you to test the model yourself in your own use case.
  • Common misconception: Some think a free local model cannot code well, but as shown, it outscores models four times larger.

    Others think setup is too technical, but it is only three terminal commands.

07Who Should Use North Mini Code and Why?

North Mini Code is built for anyone who codes and wants a fast, private, and free AI assistant that runs on their own machine. It is especially useful for:

  • Developers who want to test ideas locally without sending code to the cloud.
  • AI agent builders who need a model that can drive tools, run terminals, and think between steps.
  • Privacy‑conscious users who do not want their data leaving their computer.
  • Hobbyists and students who want to experiment with local coding models without paying for subscriptions.
Because it is Apache 2.0 licensed, you can even use it in commercial products. In this episode, Julian showed that a 30‑billion‑parameter model can outperform giants four times its size, making powerful AI accessible to everyone.