01What is Qwable 5 27B and How Does It Work?

Qwable 5 27B is a free, open-source coding model that runs locally on your computer. It is designed to help you build websites, games, and other software without needing an internet connection or paid API. The model is a fine-tuned version of the Qwen 3.6 27B base model from Alibaba, giving it a 'coding jacket' that makes it better at programming tasks.

In this episode, Julian Goldie showed how he used Qwable 5 27B to create a landing page and a 3D dragon game. The results were surprisingly good for a local model. He noted that the quality felt like 'last year's frontier'—not quite as good as top paid models like Opus 4.8, but much better than most local models he has tested.

The model runs locally using Apple MLX, a free open-source setup for running large language models on Macs. You cannot run it with Ollama, so you need to use MLX or a similar system.

02Main Features of Qwable 5 27B

Qwable 5 27B comes with several standout features that make it a strong choice for local coding. Here are the key ones:

  • Free and open source – You can download and run it without paying anything.
  • Local execution – It runs entirely on your own hardware, so no internet is needed.
  • Based on Qwen 3.6 27B – A powerful base model from Alibaba, with a fine-tune for coding.
  • Works with Apple MLX – Easy to run on Macs using this free tool.
  • Top of the leaderboard – In Julian's GoldieBench tests, it outperformed other local models like Gemma 4 12B and Qwifier 9B.
  • Integration with AgentOS – You can plug it into your agent operating system to generate live previews and save all your builds.

These features make it a practical option for developers who want to experiment with local AI without high costs.

03How to Set Up Qwable 5 27B on Your Mac

Setting up Qwable 5 27B is straightforward if you follow these steps. Julian used a Mac Studio with an Apple M4 Max and 36GB of memory, but you can adapt this to your own setup.

  1. Download Qwable 5 27B from Hugging Face. The model is available for free there.
  2. Install Apple MLX, which is a free open-source tool for running LLMs on Macs.
  3. Load the model into MLX using the Hugging Face integration.
  4. If you use AgentOS, plug the model into your local engine so you can use it alongside other models.
  5. Start generating code by giving it commands. You can preview your builds in a workspace.

Note that you cannot use Ollama to run this model. Stick with MLX or similar tools that support it.

04Using Qwable 5 27B to Build Projects

Julian demonstrated how to use Qwable 5 27B for real projects. He created a landing page with a moving background that looked clean and professional. He also built a 3D dragon game that worked smoothly.

To use the model, you give it a command in plain English. For example, you might say 'build a landing page' or 'create a 3D game.' The model generates the code, and if you are using AgentOS, you get a live preview right away.

Inside the AgentOS workspace, you can open your creation in a new tab or grab the code directly. Everything you build is saved, so you can come back to it later. This makes it easy to iterate and improve your projects.

Julian noted that the quality of the output was better than what he usually gets with other local models. The landing page did not look like 'generic AI slop'—it was polished and usable.

05Performance and Comparisons with Other Local Models

Julian put Qwable 5 27B to the test against other local models. In his GoldieBench testing, Qwable 5 27B came out on top.

  • Gemma 4 12B – Totally failed on the same task. The game it produced did not work at all.
  • Qwifier 9B – Also failed the same prompt. The output was not functional.
  • Qwable 5 27B – Built a working game and a nice landing page.

In terms of speed, Qwable 5 27B is slower than Qwifier 9B. Julian noticed that his Mac Studio ran a bit slower when using Qwable, but it could still run in the background while he did other tasks, such as using Cloud Desktop.

Overall, the quality of what Qwable produced was far superior to the other local models he tested. It is now his favorite local model so far.

06Tips, Limits, and Mistakes to Avoid

When using Qwable 5 27B, there are a few things to keep in mind to get the best results.

  • Expect slower performance – It is a 27B model, so it will slow down your system somewhat. Be patient.
  • Do not expect frontier-level output – It is not as good as Opus 4.8 or Fable 5, but it is impressive for a local model.
  • Use Apple MLX – Do not try to run it with Ollama, as it will not work.
  • Check your memory – Julian used 36GB of memory, so make sure your machine can handle the model.
  • Leverage AgentOS integration – Use live previews to see what you build and save everything for later.

Avoid comparing it to paid frontier models—it will not match them. Instead, focus on what it can do for free and locally.

07Who Should Use Qwable 5 27B and Why It Matters

Qwable 5 27B is perfect for developers, hobbyists, and businesses that want to experiment with AI coding without spending money on APIs. It is also great for anyone who values privacy and wants to run models entirely on their own machine.

Because it is free and open source, it lowers the barrier to entry for learning AI development. You can build real projects like websites and games, and the integration with AgentOS makes the workflow smooth.

Local AI is moving fast, and models like Qwable 5 27B are proof. Julian mentioned that new open-source models are coming out almost daily, with many updates from China. This trend means better and better local tools are on the way.

If you want to stay ahead of the curve, trying Qwable 5 27B is a smart move. It gives you a taste of what local AI can do today, and it is only going to improve.