DeepSeek minus the spying

Representation pic

For those who have been oblivious to the hype and misinformation around DeepSeek, here is a small primer. DeepSeek R1 is a large language model (LLM), much like Open AI's ChatGPT, except it is actually open source. You can download it, fine-tune it, and use it in any way you see fit. It has consistently beaten Open AI in various tests and made the model with a fraction of the resources available to Open AI.

However, DeepSeek is allegedly closely related to the Chinese Communist Party (CCP), which is why a lot of American media is sounding alarm bells. Not to mention, the low cost of training and improved performance have scared American companies. Some American politicians have also proposed banning downloading the Rs 1 model in the US, a hilariously incompetent solution.

As for OpenAI's complaint that DeepSeek uses data collected from their company to train Rs 1, Open AI stole all our data first to make their model, so I am okay with that bit of theft. If you want to know more about DeepSeek Rs 1, the company has a nice scientific paper and webpage explaining everything. Visit this link: https://bit.ly/deepsmd.

Why not use DeepSeek directly?

DeepSeek can be easily tested on its website or using its app, which is currently ahead of ChatGPT. Using the service requires registration. However, you might not want to do this because the DeepSeek company is allegedly related to the CCP. This means that data entered on the app and website might not be secure and could be shared with the current government in China. If you don't really care about data security, you can still use the website, but if you need to ensure the data is safe, this might not be the best site. So what are the alternatives?

Ollama

Ollama is for those who don't mind using text to set things up. Ollama has very easy-to-follow instructions on its website, and it also has a web interface rather than a program to run your LLM. So it feels like ChatGPT but runs only locally. Hardware restrictions apply; you can't run a full version of DeepSeek but rather a quantised version, which depends on your hardware. The list of available models is on the Ollama website.
https://ollama.com/

Groq

No, this isn't Elon Musk's AI. Instead, it is a company that specialises in hardware and software that accelerates output from LLMs. You will have to register, but once you do, you can access a distilled or reduced version of DeepSeek Rs 1, which is not only capable but also very fast. Groq can be tested for free. However, they have a pricing module if you want to use it regularly or as part of a program you have created. Given the speed, the pricing is pretty fair.

LM Studio

If you don't want a virtually hosted DeepSeek R1 and would prefer to manage your own security and data, LM Studio is a great option for running DeepSeek. That said, there are a few restrictions; the first is you will not be able to run the largest model of DeepSeek, but a much smaller model. LM Studio's interface will tell you what you can run depending on your computer's hardware. The second problem is you need a good Nvidia GPU if you want to squeeze out some quick responses. LM Studio supports running LLMs on the CPU, but the performance usually involves entering the query and grabbing a coffee while you wait. LM studio is an
entirely graphical user interface for LLMs that is easy to set up and use.
https://lmstudio.ai/

What to choose?

In most cases, running your own personal DeepSeek R1 will be restricted to the hardware you own. You can upgrade your hardware to match the specs needed to run DeepSeek or any other LLM, but this usually means sinking a lot of money into hardware. In most cases, paying off a service like Groq and using the LLM on their site would be better. Suppose you choose to host your own; by May, the new Jetson Nano Super single-board computer will be released in India, and that is capable of running a smaller version of DeepSeek without essentially breaking the bank. Hosting your own also lets you switch seamlessly between models, letting you use fine-tuned purpose-built LLM models like Codestral on the same hardware setup. So, my choice for immediate and fast but safe use of DeepSeek R1 is Groq, but I recommend self-hosting for the long-term and ultra-secure setup.