This post has been updated: As of Febuary 6th 2025 we have expanded our on device models to include the three more on device models:
- Llama 3.2 1B – Lightweight, good for multilingual chat.
- Llama 3.2 3B – More capable, handles complex multilingual tasks.
- DeepSeek R1 7B – Balanced power, great for advanced reasoning.This post was updated on February 06, 2025
Two years ago, we launched our first AI Assistant for Craft. It was powered by OpenAI’s cloud-based models, offering powerful text generation and reasoning capabilities. We were one of the first productivity tools to add GPT-3, but it required an internet connection. Since then, we’ve watched AI evolve rapidly and we’ve listened to our community’s desire for both privacy and speed, without always needing a connection to the internet.
At the end of last year we took our first step in the direction of on device AI by enabling Apple Writting tools and image playground features. Today, we’re taking a step forward by adding the smallest DeepSeek R1 1.5B model into our AI Assistant. R1 is an open-source Large Language Model (LLM) that you can install locally and run in Craft .
Privacy, Data Security, and Content Control
Please note that this feature is experimental. When you first download or update Craft, DeepSeek R1 is not automatically installed or enabled. The model is still a work in progress and may make mistakes. You can download and install DeepSeek R1 1.5B from your account settings within Craft. Because the model runs on your device, none of your data is sent over the internet, so you maintain full control and privacy.If you choose to download and enable R1, your data stays on your device. This makes it perfect for those everyday tasks where convenience matter most—like taking quick notes, organizing to-dos, or summarizing shorter pieces of text. It also allows you to use basic reasoning capabilities without an internet connection.
If you decide the model isn’t for you, nothing will be installed on your device, if you tried it out, but decided to not use it anymore you can easily remove it from your app, and it will be deleted from your device completely.
On-Device AI vs. Traditional AI Models
Most AI models from companies like OpenAI and Google require a server connection. Every time you ask a question, your request is sent securely to a server for processing. In contrast, on-device models use only your device’s memory and processing, and never send your data to the cloud. While traditional models are powerful and continuously improving, on-device models are still in their early stages. However, they offer complete privacy and the convenience of working offline, helping you stay in your flow.
On device AI models work a little bit like electric cars. They deliver instant acceleration, don’t require regular stops at the gas station, and they are ideal for shorter commutes. For longer journeys across varied terrain, however, a traditional gas-powered engine may still be more reliable. Interestingly, some of the most high-performance cars in the world are hybrids – combining electric motors with gas engines for the best of both worlds.
We believe AI follows a similar path. On-device models like the smaller R1’s or Apple Intelligence can handle quick, day-to-day tasks, while advanced, cloud-based models like ChatGPT are still best for heavy lifting and deep reasoning. Our vision is to offer both, letting you pick and seamlessly transition between “engines” within Craft’s AI Assistant.
From the beginning, our mission has been to create a flexible, offline-first space that puts you in control. Our integration with DeepSeek R1 is an early but important experiment, pointing to a future where our devices become increasingly intelligent. While we keep working to ensure that Craft remains the flexible, fast, and private home for all your best ideas.
Let us know what you think and don't worry, #tags are still coming.
Balint
