You can now use the open source DeepSeek R1 model locally with Opera Developer

February 17th, 2025

We’ve recently enabled support for the open source DeepSeek AI model in Opera Developer. Last year, we made it the first web browser with built-in local LLMs, giving you access to over 50 families of LLMs. Opera Developer utilizes the Ollama framework (implemented by llama.cpp), which now enables the DeepSeek R1 model family – this means that you can run the model locally and without any data leaving your device.

DeepSeek R1 is an open source reasoning model that includes an LLM, and if you want to try to run it locally on your device you’ll need to download it. With Opera Developer you’re just a few clicks away from privately and securely running R1 in your own machine – let’s get into how you can do this.

Accessing DeepSeek R1 models in Opera Developer

To access the various DeepSeek R1 models you first need to download Opera Developer – or update it to the latest version if you already have it. Once you’ve got it, open the browser and click on the Aria logo located to the left – in the sidebar. Once inside the Aria interface, expand the chat history by clicking this logo in the top left corner:

Then access the settings menu –that’s the big button underneath the chat history list. Once in the settings menu, click on the “Local AI Models” option:

In the search field, type in “deepseek” and select one of the available models to download.

Click on the download button and that’s it – you’re all set! All you need to do now is to open a new chat with Aria and choose the model you just downloaded.

You might have noticed that there are various DeepSeek R1 models to choose from, and the reason for that is for you to be able to choose one that your computer can run. Generally, we would recommend you to try out a smaller model, like the deepseek-r1:14b (8.4GB). If you’re unsure of your computer’s local AI capabilities, we’ve got you covered: we’ve also made a benchmarking tool that helps you assess this.

Why do reasoning models matter?

Reasoning models like DeepSeek R1 or Open AI’s o1 and o3 perform better in certain benchmarks than regular LLMs. This is because they’re reasoning models that, once you give them a prompt, will go over several iterations of the output to verify the quality of their answer before providing it to you.

Because of the way in which these reasoning models approach your prompts, you can expect answering times to be significantly higher – at least in comparison to the answering times of what we currently call “normal LLMs”. One of the interesting things about DeepSeek R1 is that you can see the train of “thought” or the chain of ideas that it is executing to reach an answer – as if it was having a conversation with itself. Whereas Open AI’s o1 and o3 are only available for paying customers, DeepSeek R1 is openly available (an open-weight model), and this is how Opera can now offer a true reasoning model that you can download and run locally, all through the Opera Developer browser.

Nowadays, you can run a rather powerful model on your machine. But we believe that access to these models shouldn’t be limited to the most tech-savvy, and that’s why we’re making sure that you can easily download and try them with Opera Developer.

Testing the capabilities of a reasoning model

If you want to try out the reasoning capabilities of DeepSeek R1 or any other reasoning models, you can ask how many “R” are in the word “Strawberry” – seriously, it gets it right now. In all seriousness, you can use one of the Math Olympiad problems as they should pose a fair test for the model. Let’s try this one out:

Is there a point inside a square with side length 20 whose distances from the sides of the square are four consecutive integers?

At this point, Deepseek showcases its “reasoning” process – its train of ideas – that lead it to solve the problem:

The answer to this problem is indeed that such a point does not exist within a square. Trying to answer this problem with a normal LLM – for example, GPT 4o – provides inconsistent results, sometimes might be right but most times it is wrong. However, reasoning models like Deepseek R1, o1 or o3 solve the problem with consistency.

Something to keep in mind when interacting with any reasoning models is that they, by their nature, take a longer time to process a query and produce an answer. Depending on your hardware and complexity of the model you are using, it may process for several minutes before you start seeing anything.