Resource-Efficient Fine-Tuning of DeepSeek-R1

Author(s): Thuwarakesh Murallie

Originally published on Towards AI.

How to make DeepSeek R1 to reason with your private data

This member-only story is on us. Upgrade to access all of Medium.

Photo by Dan Schiumarini on Unsplash

We no longer seek validation to say that DeepSeek R1 is awesome. It really is.

DeepSeek can provide comparable performance to OpenAI-O1 but at a fraction of its cost. That’s because the model is open source.

Yet, we cannot jump in and start using it in production-grade applications. Any model, open-weight or proprietary, has an inference cost that must fit your budget. Needless to say, the more parameters there are, the more costs there are.

DeepSeek’s R1 is a 671 billion parameter model. For any given inference, some 37 billion active parameters will do the job for you. You need at least 90GB of vRAM to host this model locally.

Here’s a back-of-the-envolop calculation of the GPU memory needed to host the model.

vRAM estimation for hosting LLMs locally — Image by the author

If you don’t have a GPU with that much vRAM, luckily, this is not the end of the world.

Of course, you can rent a GPU from a provider. But I’m not talking about that.

DeepSeek R1 can be successfully run on the most popular GPUs, such as RTX 3090 or 4090. You can even run it for… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI