I wouldn’t say DeepSeek is outright better. It outperforms in certain benchmarks, but it still has limitations. Think of it like the evolution from old mainframe computers to early personal computers. When personal computers first came out, they were nowhere near as capable as mainframes—they were just “good enough” for some tasks. DeepSeek is similar: it’s good enough for certain applications, but it’s still not as reliable as the larger models, especially when it comes to mathematics.
Don’t get me wrong—DeepSeek is a game changer, particularly because you can run the model locally, assuming you have the necessary resources. There are smaller versions that can even run on a Raspberry Pi, but they’re not very accurate and tend to be extremely slow unless you have a separate GPU. The largest model, on the other hand, is far more accurate, but running it locally requires substantial hardware. You’d need about half a terabyte of storage, not to mention significant processing power. It’s possible, but definitely not practical for the average person.
The largest DeepSeek model has around 671 billion parameters. To run it, you’ll need a GPU with at least 40+ GB of VRAM, 128 GB of RAM, a powerful CPU, and a large amount of storage. They do offer smaller models that can run on less powerful hardware, and you can also access the R1 version online, similar to ChatGPT. However, I wouldn’t recommend that option—everything you input is transmitted directly to the Chinese government. ChatGPT’s data privacy concerns are already unsettling enough, and I can’t imagine what the CCP might be doing with that kind of data.
Running the model locally helps you avoid this security risk, but be prepared for the cost. To build a PC capable of running the DeepSeek-R1-Zero model locally, you’re looking at an estimated cost of around $12,200 to $19,000 USD, depending on the specific components you choose.