DeepSeek- R1 and FP8 Training: A Breakthrough in Model Efficiency

By Ashni Singh,

Choate Rosemary Hall, CT

On the 16^th of February, China’s new AI startup DeepSeek released its R-1 model, taking the world by a storm and giving Open AI’s model a run for its money across dimensions like coding, understanding of the English and Chinese language and cost. R1, a “reasoning” model, claims to be more efficient and was built at a fraction of the cost of proprietary models like Open AI’s o1 model or Meta’s Llama 3 model. This is no small feat and an important milestone for the open-source community.

So what does all this mean? In this article we will unpack various key elements including implications on the training of current AI models, potential use cases and the impact, if any, on tech giants like NVIDIA and OpenAI.

DeepSeek’s technical report goes into a lot of detail on various performance optimization techniques that enabled their breakthrough results on efficient LLM training. The most important of these is the FP8 mixed precision training strategy. So what are some key benefits and implications of the FP8 framework?

Greater Efficiency, Scalability and Time Savings: Leveraging FP8 precision means reducing the memory footprint on models. This means that larger models can be trained on the same hardware resources. Additionally, with MLA and FP8 kernel optimizations, DeepSeek models see massive improvements in throughput and memory capacity. Extrapolating from this, FP8 could allow for training more complex models or larger batches at once allowing for faster research and development. This could also lead to faster model training times meaning that more experimentations can be run in the same amount of time. It is also significantly cheaper, and the inference cost (the cost it takes to operate and generate a response) is much lower.

Democratization of AI: Due to its significantly lower cost (it is 96% cheaper than GPT-40), DeepSeek challenges the notion that only major tech companies with huge budgets can utilize and operate AI. This indicates major shifts in the tech and AI industries, and means that DeepSeek is pioneering accessible AI technology. This is especially due to its open-source accessibility, which means that anyone is able to customize DeepSeek models.

Image of Computer Chip

Potential Use Cases

The most important use case or beneficiary of FP8 will be LLM models like GPT which are used in Natural Language Processing (NLP). Using this new technology organizations will be able to build better Natural Language Processing (NLP) models that can handle complex queries, generate human-like responses and make better predictions.

Additionally, DeepSeek models have been shown to rival and even surpass models like GPT-40 when it comes to accuracy as a coding assistant, a tool for education, or even a multilingual system, as it excels in the Chinese language.

Healthcare and drug discovery could be another transformative and powerful use case. AI models used in healthcare require tremendous amounts of data and corresponding computational power.  The ability to train large models more efficiently could lead to strides in drug discovery and patient diagnostics.

Machine Learning Model Architecture with Multiple layers

The autonomous vehicle market will also benefit from this. The training of AI models for autonomous vehicles is extremely complex as it requires a large amount of data sets and real time processing. These systems daepend heavily on machine learning for image recognition and real time decision making. Reducing the computational cost of training these systems as well as making it faster to develop and deploy could lead to major advancements in self-driving car technology.

In short, FP8 can lead to a game changing shift in AI model development for industries across the board from climate and environmental sciences to gaming. Its potential is immense and still untapped. As this technology continues to evolve it will make AI more accessible, scalable and sustainable than ever before.

----Works Cited

Duberstein, B. (2025, January 27). China’s DeepSeek AI Model Shocks the World: Should You Sell Your Nvidia Stock? Retrieved from The Motley Fool website: https://www.fool.com/investing/2025/01/27/chinas-deepseek-ai-model-shocks-world-sell-nvidia/

DeepSeek-R1 and FP8 Mixed-Precision Training. (2025, January 27). Retrieved February 17, 2025, from Colfax Research website: https://research.colfax-intl.com/deepseek-r1-and-fp8-mixed-precision-training/

Gupta, P., & Kiely, P. (2024, March 14). 33% faster LLM inference with FP8 quantization. Retrieved February 17, 2025, from Baseten website: https://www.baseten.co/blog/33-faster-llm-inference-with-fp8-quantization/

Enhancing DeepSeek Models with MLA and FP8 Optimizations in vLLM - Neural Magic. (2025, February 4). Retrieved February 17, 2025, from Neural Magic - Software-Delivered AI website: https://neuralmagic.com/blog/enhancing-deepseek-models-with-mla-and-fp8-optimizations-in-vllm/

Gupta, D. (2025, January 25). DeepSeek AI: Revolutionizing Efficiency, Innovation & Affordability in Next-Gen AI. Retrieved February 17, 2025, from Deepak Gupta | AI & Cybersecurity Innovation Leader | Founder’s Journey from Code to Scale website: https://guptadeepak.com/deepseek-revolutionizing-ai-with-efficiency-innovation-and-affordability/

DeepSeek-AI, Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., … Luo, F. (2024). DeepSeek-V3 Technical Report. Retrieved from arXiv.org website: https://arxiv.org/abs/2412.19437