Understanding the Capabilities of DeepSeek R1 Large Language Models

DeepSeek R1 has, for good reason, taken the AI/ML community by storm these past weeks, and has even in fact spread beyond to the wider world with major effects on both the economy and politics. This is largely because of the model suite’s open-source nature & incredibly low training price, which has shown the greater community that training SOTA AI models my not require nearly as much capital or proprietary research as previously thought.

In the first part of this series, we introduced DeepSeek R1 and showed how to run the model using Ollama. In this follow up, we will begin with a deeper dive into what actually makes R1 so special. We will focus on analyzing model’s unique Reinforcement Learning (RL) paradigm to see how reasoning capabilities of LLMs can be incentivized purely through RL, and, afterwards, discuss how the distillation of these techniques to other models allows us to share these capabilites with existing releases. We will conclude with a short demonstration on how to setup and run DeepSeek R1 models with GPU Droplets using 1-Click Model GPU Droplets.