Author: GPT4All Team

Run DeepSeek R1 Privately and Locally with GPT4All

We’re excited to introduce open-source support for DeepSeek R1 distillations in GPT4All. This release allows anyone to run DeepSeek R1 distillations locally and privately on everyday hardware.

What is DeepSeek?

DeepSeek R1 is an AI model designed to reason through its responses before providing answers. It significantly improves response quality compared to previous language models.

What is GPT4All?

GPT4All is the open-source local LLM application. It requires zero coding and AI experience to get started: anyone can download the application and run LLMs locally and privately on their own devices.

How Does GPT4All Support DeepSeek R1?

• The 'Thinking' Interface: DeepSeek R1 generates reasoning traces using "thinking" tokens. The GPT4All desktop application introduces a collapsible Thinking bar, allowing you to inspect the reasoning trace or hide it as needed.
• Support for R1 Distillations: Running the full DeepSeek model (640B parameters) on everyday laptops and desktops is not currently feasible. However, distilled versions of R1 adapted into the Qwen and LLaMA architectures are supported in GPT4All. While these versions exhibit some quality degradation compared to the full-parameter model, they still outperform comparable models with the same number of parameters.

Getting Started with On-Device Reasoning Models Like DeepSeek

To start using DeepSeek R1 with GPT4All:

1. Install the GPT4All App: Download the latest version of the app from our official site.
2. Go to Models: Find the DeepSeek model in the Recommended Models section.

Chat with your private data with DeepSeek

You can use GPT4All LocalDocs and let DeepSeek access your computers file system. This enables applications like reasoning powered, private data chat.

Explore DeepSeek Reasoning Data

Reasoning models output long traces of text in their responses, "thinking out loud" as they explore the space of relevant information on the way towards finding an answer to complicated problems.

This data map, powered by the LLM Observability Connector in Nomic Atlas, lets you explore 250,000 of DeepSeek R1 reasoning traces organized into regions of similar content:

The data map is interactive: you can zoom in and out to explore at different levels and hover over a point to show its data.