Introduction to DeepSeek-R1
The major news this week is the release of DeepSeek-R1, which is the first generation reasoning model from Deepseek (an AI startup based of China). (For a more detailed background of Deepseek, please read our post on DeepSeek-v3 from last week).
DeepSeek-R1 has been hailed by tech luminaries throughout this week, with Mark Andreessen calling it “One of the most amazing and impressive breakthroughs I’ve ever seen — and as open source, a profound gift to the world.”
DeepSeek-R1 is comparable in performance to OpenAI’s reasoning model o1. It actually beats o1 in 5 of 11 benchmarks, and comes close to o1 in the remaining six. It also handily beats Anthropic’s Claude 3.5 Sonnet and OpenAI GPT-4o on almost all tested benchmarks. Benchmarks aside, the chatter on X reveals that people are just loving Deepseek-R1. There is a viral TeamBlind post showcases Meta’s GenAI team (responsible for the Llama models) in panic mode seeing Llama4 behind the latest Deepseek models.
This post will not discuss any of the technical details behind Deepseek-R1, there is already a lot of literature at Interconnects, in The Batch, besides Deepseek’s git, and their extremely well written technical paper. But the short summary is that the model has been trained at a very low cost (Estimates from the paper suggest a $5.5 Million Compute budget), trained without using fancy GPUs (Deepseek uses a cluster of 2048 H800 GPUs), and truly showcases the power of Reinforcement Learning (RL) when used in combination with other methods such as fine-tuning on Chain of Thought (CoT) examples.
Trying out DeepSeek
This post will focus on our experience of getting DeepSeek-R1 to work with our beloved Vector robot. But first let’s look at the ways by which you can try DeepSeek-R1 yourself. Here are some known venues:
chat.deepseek.com Deepseek offers their model on their portal and also via API. You need a Google account to access their portal, or alternatively you can create an account with your email address. Mainland China users can use their phone number for logging in.
One of the cool things of DeepSeek R1 is that unlike OpenAI o1, DeepSeek R1 shows you what it is thinking about. The following screenshot shows that DeepSeek has reasoned out the answer for 26 seconds. It shows the logic it used for reasoning.
Now, after reasoning, the concrete answer is delivered. In the following screenshot, the text in light gray is the reasoning part, while the text in dark is the actual answer reasoned out by the model. You could ignore the think part if you are interested solely in the answer, or you could use the think part to determine what logic the model used to reason.
chat.deepseek.com is certainly as great way to get a flavor of the model.
You could also install the DeepSeek App on your phone which will connect to the DeepSeek portal on the backend. As of today, the DeepSeek App is the number 1 downloaded app on Apple App Store, ahead of ChatGPT App from OpenAI.
Together.ai and Fireworks.ai: Deepseek opensourced DeepSeek-R1 on HuggingFace. Both Together and Fireworks offer DeepSeek-R1 on their Playground and via API. DeepSeek-R1 is the default LLM on the Together.ai chat interface. Note that API access at both Together.ai and Fireworks.ai is fairly costly ($7 and $8 per 1 Million Output Tokens respectively). API access at Deepseek is at $2.19 per 1 Million Output Tokens. But one crucial difference is whether you wish to access the model at an American data center or one in mainland China.
Given the popularity of DeepSeek-R1, it will surely be available widely. As an example, the CEO of Perplexity also announced their intention to host DeepSeek R1.