AI models

The open source models WorkOffline runs on your laptop. Each one downloads once and then works with no internet, and because it runs on your laptop it's private: your questions and files are never sent to the model's maker or anyone else. Single and Premium get the recommended model for your laptop; Premium adds the rest.

Everyday: Lite, Standard and Large

Chat with your files, writing, summaries and questions. The app recommends one size for your laptop and downloads it at setup. The other everyday models are in the Premium model library.

Large needs 32 GB of memory and a Mac with an Apple chip or a graphics card with 12 GB or more. Other 32 GB computers get Standard, because Large answers slowly there. How the app picks.

Coding

Coding models write the small tools behind Build mode. Pick the size that suits your memory.

Which coding model, and what Build mode can do.

Pictures and scans

These models read photos, screenshots and scanned pages as well as words.

Comes with every laptop, alongside your everyday model.

Coming to Premium Next update

Google, Meta, OpenAI and NVIDIA models join the Premium library in the next update, once they pass our tests.

The makers

  • Alibaba's Qwen team

    The Qwen team at Alibaba Cloud publishes open models in many sizes. Qwen3 (2025) gives us Lite, Standard and Large, Qwen3 VL reads pictures, and Qwen2.5 Coder writes code. Qwen2.5 Coder comes in six sizes from 0.5 to 32 billion parameters and was trained on 5.5 trillion tokens of code and text, according to its model card. All the Qwen models here are Apache 2.0.

    Qwen3 Technical Report

  • Microsoft

    Phi models come from Microsoft Research. They are trained with a lot of carefully made teaching material and a focus on reasoning, which makes them good at maths, logic and well structured answers for their size. MIT licence.

    Phi-4 model card

  • Hugging Face

    SmolLM3 is made in the open by Hugging Face, with its training recipe published for anyone to read. It is the smallest chat model here. Apache 2.0.

    SmolLM3 model card

  • IBM

    Granite is IBM's family of open models, built with business use in mind. Apache 2.0.

    Granite 3.3 8B Instruct model card

  • Mistral AI

    Mistral AI is based in Paris. Mistral Small 3.2 is the biggest model in our library and can read photos, screenshots and scans as well as write. Apache 2.0.

    Mistral Small 3.2 model card

  • Nomic

    Nomic builds open tools for search. Nomic Embed does not chat: it turns passages from your files into something the app can search by meaning. Apache 2.0.

    nomic-embed-text-v1.5 model card

  • Google

    Gemma is the open model family from Google DeepMind, built from the same research as Gemini. Gemma 4 comes in sizes from small laptop models to a 31 billion parameter model, and reads pictures as well as text. Google publishes official 4-bit versions made for computers like yours. Apache 2.0. Coming to Premium.

    Gemma 4 E4B model card

  • Meta

    Llama is Meta's open model family and one of the most widely used. Llama 3.1 8B and 3.2 3B are the sizes that suit a laptop. They come under Meta's Llama Community License, which asks for a "Built with Llama" notice and that you follow Meta's usage policy. Coming to Premium.

    Llama 3.1 8B Instruct model card

  • OpenAI

    OpenAI, the maker of ChatGPT, publishes two open models. gpt-oss-20b is the one that fits a laptop with 16 GB of memory or more; the bigger gpt-oss-120b needs a workstation. Apache 2.0. Coming to Premium.

    gpt-oss-20b model card

  • NVIDIA

    Nemotron is the open model family from NVIDIA. Nemotron 3 Nano 4B is a small model trained for following instructions and using tools, and NVIDIA publishes a ready-to-run version for laptops. NVIDIA Open Model License. Coming to Premium.

    Nemotron 3 Nano 4B model card

How we choose models

  • Licences that allow commercial use and redistribution. Non-commercial and research-only models are not offered. The app shows each model's licence, maker and any required notice (such as "Built with Llama") before you download it.
  • Pinned and checked. Models download from our own Hugging Face organisation, WorkOffline, pinned to an exact version and checked against a SHA-256 fingerprint before use.
  • Tested on real work. Candidates go through our bake-off: about 190 questions over sample work files, scored for correct answers, citations, saying so when nothing is found, speed and memory use.

Your own model Premium

Load a GGUF model file from your computer, or download one from any Hugging Face repo you choose, at your own responsibility. Before first use the app shows what the file says about itself, asks you to agree once, and checks that it runs, fits your memory and is intact. It is labelled Your own model and can be used for chat or for a single task.

See Premium