Which local AI model should I use for coding, and what can Build mode do?
Pick by memory. With 8 GB, Qwen2.5 Coder 1.5B Instruct (about 1.1 GB). With 16 GB, Qwen2.5 Coder 7B Instruct (about 4.7 GB). With 32 GB or more, Qwen2.5 Coder 14B Instruct (about 9 GB). In WorkOffline these models power Build mode (Premium): you describe a small tool, the model writes it, and it runs on your laptop with no network, showing you what it would make before anything is saved. They are good at short, well described jobs on your files, and much smaller than cloud coding assistants, so check their work.
What a coding model is
A coding model is a language model trained mostly on source code and on text about code. WorkOffline offers the Qwen2.5 Coder family from Alibaba's Qwen team. Its model card says the family was trained on 5.5 trillion tokens of code, text and synthetic data, comes in six sizes from 0.5 to 32 billion parameters, and is released under Apache 2.0. We offer three of those sizes, chosen to fit normal laptops.
Which model for your laptop
| In the app | Model | Parameters | Download | Memory |
|---|---|---|---|---|
| Coder Lite | Qwen2.5 Coder 1.5B Instruct | 1.54 billion | about 1.1 GB | 8 GB or more |
| Coder | Qwen2.5 Coder 7B Instruct | 7.61 billion | about 4.7 GB | 16 GB or more |
| Coder Large | Qwen2.5 Coder 14B Instruct | 14.7 billion | about 9 GB | 32 GB or more |
- 8 GB: Coder Lite. Fine for short scripts: listing, sorting and renaming. Keep the request small.
- 16 or 24 GB: Coder, for most Build mode jobs.
- 32 GB or more: Coder Large. Writes longer, more careful tools. Like the Large everyday model, it is quickest on a Mac with an Apple chip or a PC with a large graphics card.
The app gives the two smaller coding models 16,384 tokens of room to work and Coder Large 32,768. The coding model is a separate download from your everyday model, so check your free disk space too.
What Build mode does
- Describe the tool. For example: "make a CSV listing every PDF with its page count and size". The coding model writes a short script. You can read and edit it.
- Pick the folders it may read. It sees nothing else.
- Try it. The script runs on your laptop in a sealed-off helper with no network, and returns a list of what it would make: new files, or renamed copies of files. Nothing is saved yet.
- Confirm. Only then does it create the new files, in a new folder in Done in your WorkOffline folder. Your originals are never changed.
Tools you keep can be opened again from Home. Build mode is a Premium feature. Repeating jobs that need no code, such as watching a folder and offering to compress new files, or summarising every file in a folder, are automations, included with Single and Premium.
Honest limits
- Small models make mistakes. They can misread a request, miss an edge case or write a script that fails. The preview step is there so you catch that before anything is saved.
- Short jobs work best. One clear task per tool. A long list of steps is better split into several tools.
- Build mode is deliberately narrow. Its tools cannot reach the internet, install anything, change or delete your files, or see folders you did not pick.
Privacy
The model, the script and your files all stay on your laptop. Build mode makes no network connections, so it works the same in a tunnel as at your desk.
Local or cloud?
Cloud coding assistants run much larger models on their providers' servers. They are better at big codebases, long programs and hard problems, and many plug straight into code editors. They need a connection, and your code is sent to the provider under its terms.
A local coding model is smaller and slower, but it works offline, costs nothing per use and keeps your code and files on your laptop. For small tools that work on your own files, that is often enough. For serious software work, many people use both.
Frequently asked questions
Can a small AI model write code offline?
Yes. Coding models of a few billion parameters run on an ordinary laptop with no internet. They handle short scripts well and struggle more with long, many-step programs.
Which coding model should I use with 16 GB of memory?
Qwen2.5 Coder 7B Instruct. It is about 4.7 GB to download and is the size meant for 16 GB laptops.
Is Build mode included with Single?
No. Build mode and the coding models are Premium features. Automations such as watch folders and batch jobs are included with Single and Premium.
Can Build mode change or delete my files?
No. A Build mode tool can only read the folders you pick, and after you confirm it can only create new files, in a new folder in Done. Originals stay as they are.
Can I use these models as a coding assistant in my editor?
Not in WorkOffline. Here they power Build mode. The models themselves are Apache 2.0, so you can also run them in other tools that load GGUF files.
Sources
- Qwen2.5 Coder 1.5B Instruct model card (Hugging Face): 1.54 billion parameters, Apache 2.0
- Qwen2.5 Coder 7B Instruct model card (Hugging Face): 7.61 billion parameters, Apache 2.0
- Qwen2.5 Coder 14B Instruct model card (Hugging Face): 14.7 billion parameters, Apache 2.0
- Qwen2.5-Coder Technical Report (arXiv 2409.12186), September 2024
- Apache License, Version 2.0 (Apache Software Foundation)
- WorkOffline AI models: sizes and licences from the app catalogue