Lite, Standard and Large: how WorkOffline picks a model for your laptop

WorkOffline runs its AI on your laptop, so the model has to fit your laptop. Rather than ask you to understand parameters and quantisation, the app offers three sizes and recommends one on first run. Here is exactly what they are and how it decides.

The three models

All three are open models from the Qwen3 family by Alibaba's Qwen team, in the 4-bit Q4_K_M format. Sizes and licences are from the publishers' model cards and files on Hugging Face (see Sources).

Bigger models write better and handle trickier questions, but need more memory and answer more slowly. The app also gives bigger models more room to work: Lite reads up to 8,192 tokens at a time (roughly a few thousand words), Standard 16,384 and Large 32,768, and bigger models take more search steps through your files before they answer.

The rules, step by step

1. Memory first. Under 12 GB: Lite. From 12 GB up to (but not including) about 28 GB, which covers 16 and 24 GB laptops: Standard.
2. 32 GB or more: Large on an Apple Silicon Mac, because its graphics chip shares that memory and runs the model quickly. On other computers, Standard, because Large would fit but answer slowly.
3. Intel Mac: the same memory rules, plus a note that answers will be slower because the model runs on the processor.
4. Disk check: if there is not enough free disk for the chosen size, plus room to spare, it steps down a size and tells you. If there is not enough even for Lite, it tells you how much to free up.

The thresholds sit between the common memory sizes on purpose. Computers report slightly less memory than it says on the box (a "16 GB" laptop often reports about 15.6), so the app does not get caught out by rounding.

Examples

The search helper

Alongside the main model, every laptop gets a small search helper, nomic-embed-text v1.5 (146 MB, Apache 2.0). It turns passages from your files into something the app can search by meaning, so "notice period" can find a clause that says "either party may terminate on three months' written notice". The main model then answers from those passages and cites the file and page.

Changing your mind

The recommendation is shown during setup with the reason, so you can see why. Every model file is downloaded once, checked against its published fingerprint (a SHA-256 hash) before use, and then works offline. Choosing and swapping models later from the full model library, and setting a different model for each task, are Pro features.

The Pro model library adds other open models alongside the three recommended ones. Each is listed with its licence from the publisher's model card:

• Phi-4 mini (Microsoft Phi-4-mini-instruct, MIT): writing and chat, 8 GB laptops.
• SmolLM3 (Hugging Face SmolLM3 3B, Apache 2.0): writing and chat, the smallest and quickest.
• Granite 8B (IBM Granite 3.3 8B Instruct, Apache 2.0): writing and chat, 16 GB laptops.
• Phi-4 (Microsoft Phi-4 (14B), MIT): writing and chat, 32 GB.
• Mistral Small (Mistral Small 3.2 24B Instruct 2506, Apache 2.0): writing, chat and pictures, 32 GB or more.

Pro also adds models for reading pictures and scans, and coding models for Build mode.

Which ticket?

Lite, Standard and Large work with every ticket, and the model download is included at setup. Pro adds the model library and model swapping, plus a vision model for pictures and scans. See tickets and prices.

Frequently asked questions

An evergreen guide from WorkOffline. No tracking scripts on this page.