DeepSeek R1 is good enough | Tigris Object Storage

DeepSeek R1 is a mixture of experts reasoning frontier AI model; it was released by DeepSeek on January 20th, 2025. Along with the model being available by DeepSeek's API, they released the model weights on HuggingFace and a paper about how they got it working.

DeepSeek R1 is a Mixture of Experts model. This means that instead of all of the model weights being trained and used at the same time, the model is broken up into 256 "experts" that each handle different aspects of the response. The main advantage of a Mixture of Experts model is that it allows you to get much better results with much less compute spent in training and at inference time.

DeepSeek R1 is a reasoning model. This means that the model generates a bunch of "reasoning tokens" before it gives you an answer. This is a very new workflow and it was originally spelled out in Learning to reason with LLMs by OpenAI in September 2024. This is an extension of Chain-of-Thought prompting as pioneered by the Google Brain team in 2022.

Finally, DeepSeek R1 is a frontier-grade model, which means that the results are really good. Specifically, it means that the results of the model are almost as good as the big closed models like Anthropic Claude, OpenAI's GPT and O line, and Google's Gemini models.

Example Output

How much wood could a woodchuck chuck if a woodchuck could chuck wood?

Thought for 45 seconds
Okay, so the question is "How much wood could a woodchuck chuck if a woodchuck could chuck wood?" ...

Classic Tongue Twister Breakdown:

1. Tautological Answer: "A woodchuck would chuck as much wood as a woodchuck could chuck if a woodchuck could chuck wood."
2. Hypothetical Estimate: Some humorous calculations suggest a woodchuck might move 700 pounds of wood or 3.5 cords of wood.
3. Literal Reality Check: Woodchucks (groundhogs) don’t actually chuck wood—they dig dirt.

Final Answer: "A woodchuck would chuck all the wood it could... if it could chuck wood! 🌲😄"

Why is DeepSeek R1 so good?

One of the big questions that people have is "if Chinese AI companies have to deal with sanctions that limit their access to GPUs, why is one of the best models from a Chinese AI company?" There are several reasons, including innovative optimizations despite hardware limitations.

DeepSeek R1 trains on a cluster of NVidia H800 GPUs, utilizing clever techniques like:

• 256 experts with specific training frameworks.
• 8-bit floating point numbers for efficiency.
• Optimized memory usage to minimize costs.

If DeepSeek had unrestricted access to advanced GPUs, many of these efficiency advancements wouldn’t have been necessary.

DeepSeek's Origins

DeepSeek is a research subsidiary of High-Flyer, focusing solely on improving AI models without commercialization plans. This structure allows them to stay a pure research organization.

The Model Weights are Open to the Public

DeepSeek R1 model weights are available for public use on Hugging Face. This allows for privacy and customization, enabling users to fine-tune the model for specific tasks.

In Summary

Overall, DeepSeek R1 is an excellent model suitable for various applications. It stands out with its openness, low cost, and capability for local deployment. Users are encouraged to experiment with this model to leverage its advanced features.