First off, DeepSeek is an AI company that develops open-source LLMs or large language models. It is based in Hangzhou, Zhejian and is owned by High-Flyer, a Chinese hedge fund. Co-founder, Liang Wenfeng established DeepSeek in December 2023, and is its CEO.

What we’re going to focus on here is the LLM itself, which is also called DeepSeek and has various versions. The main version in the spotlight right now is the one built on the R1 LLM.

R1 is DeepSeek’s open-source reasoning model. On its site, it claims that R1 has performance on par with OpenAI-o1, which is another LLM that’s designed to solve complex problems.

R1 stands out from the crowd as it has around 670 billion parameters. Parameters are essentially ‘variables’ in an AI model, that determine how it processes data and the following results. It’s not always the case that more parameters equals a better working AI though. According to ourworldindata.org, parameters can assist in higher accuracy, but having too many can run the risk of the AI learning specific examples rather that actually learning the underlying patterns. So in short, too many parameters could be detrimental to the model itself – so a balance needs to be struck.

DeepSeek functions in a very similar way to ChatGPT. Ask it questions, put in prompts, and await the results. You can download it through the Apple Apps store or on your browser via the DeepSeek site.

What’s different about DeepSeek?

Asides from being seen as the largest open-source large language model yet, DeepSeek has shaken things up by keeping costs down.

For context, building AI models is super expensive as there’s a huge cost associated with the computational power required to train the AI itself. To do this, you need to throw money at things like super-powered GPUs (like ones from Nvidia for instance), be able to accommodate enormous amounts of data, and of course have people who are specialists in this area – again, another big cost.

According to reports, the DeepSeek Ai model costs around $6 million to build. In comparison, some other models cost billions in the US.

How did DeepSeek keeps costs down?

The general consensus on how DeepSeek achieved this, as expanded on by The Next Platform, is that they massively reduced the amount of hardware needed to train and produce the AI model. Let’s break this down as simply as possible.

According to reports, DeepSeek was able to take a few thousand ‘crippled “Hopper” H800 GPU accelerators from Nvidia, which have some of their performance capped’ and create a MoE foundation model. These cards are high-performance GPUs designed for AI and machine learning. A MoE foundation model is what’s called a ‘Mixture of Experts’, where you basically have ‘sub-networks’ or ‘experts’ that are selected based on the input.

What that means is that DeepSeek achieved similar results to the likes of OpenAi, Anthropic, and Google, but went about it differently – namely by reducing the amount of hardware and resource needed to train the model, and by going with a seemingly more efficient architecture – MoE.

But again – why is this special? Well the real win here is that DeepSeek has seemingly demonstrated that you can spend a fraction of the cost to create a LLM, and you may not need as much hardware as you thought. When you look at it like that, you can see why Nvidia value dropped so suddenly, as companies may not need to buy loads of GPUs to create an AI model.

Censorship and limitations

It has been observed by some sources that the API version of R1 uses some censorship mechanics, largely around sensitive, political topics. For instance, a question prompted by the BBC regarding what happened at Tiananmen Square on June 4th 1989 returned an answer of “I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses.”

Of course, already there are examples of those looking to workaround these apparent mechanisms, and we’d predict this will continue as people’s curiosity into both the potential and limitations of R1 grows.

Final word

We have a feeling DeepSeek will be the one to watch, and we’re curious to see how Western companies respond to a new, cheaper way of building LLMs. We’ll be updating this page with more information as it comes in, so stay tuned.

About the Author

Ramzi Musa

Ramzi Musa is a Growth Manager and Editor for BGFG. He has experience and expertise in tech, gaming, and hardware.