Apr 11, 2025

DeepSeek

DeepSeek, a Chinese AI company founded in 2023 and based in Hangzhou, has made significant strides in the LLM space with its cost-effective and high-performing models. DeepSeek has gained global attention for models like DeepSeek-V3 and DeepSeek-R1, which are open-source and compete with industry leaders like OpenAI. Their approach, highlighted by a training cost of just $6 million for DeepSeek-V3 compared to OpenAI’s $100 million 💸 for GPT-4, has disrupted traditional AI development paradigms.

This guide will help you understand these models, and how to choose the best fit for your AI-powered projects.

We focus here on the latest models, but if you are interested in a broader overview, you can see this What is DeepSeek? AI Model Basics Explained video.

DeepSeek-V3: general-purpose powerhouse

DeepSeek-V3 is designed for a wide range of tasks, using a Mixture-of-Experts (MoE) architecture with 671 billion parameters, of which 37 billion are active per token.

This model excels in code generation, mathematical reasoning, and natural language processing, outperforming models like GPT-4o on various benchmarks. A significant update, DeepSeek-V3 0324, released in March 2025, enhances reasoning capabilities, achieving even better results. This version, with 685 billion parameters, is MIT-licensed and available on Hugging Face DeepSeek-V3-0324 Hugging Face, offering local deployment options, such as running on a 512 GB MacBook M3 Ultra at 20 tokens/sec via 4-bit quantization.

DeepSeek-V3 is accessed via the deepseek-chat API, it’s ideal for developers building chatbots, content generators, or coding assistants, with very approachable pricing.

DeepSeek-R1: reasoning specialist

DeepSeek-R1, released on January 20, 2025, is a reasoning-centric model built upon DeepSeek-V3, focusing on logical inference, mathematical problem-solving, and reflection. It uses a 671 billion parameter MoE architecture, with 37 billion activated per forward pass, and is trained using a hybrid approach involving reinforcement learning (RL) and supervised fine-tuning (SFT).

DeepSeek-R1’s ability to generate Chain of Thought (CoT) explanations makes it suitable for educational tools and research applications. It performs comparably to OpenAI’s o1 on benchmarks.

Accessed via deepseek-reasoner, it offers a really great pricing as for reasoning model.

Comparison

Distilled models

To cater to resource-constrained environments, DeepSeek provides distilled versions of DeepSeek-R1, with parameter sizes ranging from 1.5B to 70B. You can check out my other article regarding models distillation.

These models, fine-tuned using 800k samples curated with DeepSeek-R1, are based on architectures like Qwen and Llama, ensuring compatibility with consumer-grade GPUs.

What is great is that these distilled models are available on Hugging Face and Ollama, making them accessible for local testing and deployment.

Here are some resources, showcasing how you can run DeepSeek-R1 on a local machines like Jetson Orin Nano Super and even Raspberry Pi.

Accessing and using DeepSeek models

As already mentioned, DeepSeek models are accessible via APIs, with deepseek-chat pointing to DeepSeek-V3 for general tasks and deepseek-reasoner to DeepSeek-R1 for reasoning, both compatible with OpenAI’s format for easy integration.

For local deployment, models are available on, but not limited to, Hugging Face and Ollama (deepseek-r1, deepseek-v3).

The simplest way to give DeepSeek a try is to use the DeepSeek Chat, which is a web-based interface similar to ChatGPT, allowing you to interact with the models without any coding.

Additionally, DeepSeek is available as a mobile app as well.

Choosing the right model

Selecting the appropriate DeepSeek model depends on your application’s requirements. For general-purpose tasks like chatbots or content generation, DeepSeek-V3 is ideal due to its versatility. For reasoning-intensive applications, such as educational tools or math solvers, DeepSeek-R1 offers superior performance. If computational resources are limited, distilled models provide a balance, with larger ones like DeepSeek-R1-Distill-Qwen-32B suitable for moderate needs.

DeepSeek’s innovative models, with their cost-efficiency and open-source ethos, are really transforming AI development. By offering a spectrum of models from general-purpose to reasoning-specialized, and distilled versions for resource constraints, DeepSeek enables us, developers, to build sophisticated AI-powered applications and continue to push boundaries, promising further advancements in the field.

干得好，DeepSeek 的工程师们！