DeepSeek's new R1 AI model will consume much less energy

DeepSeek, which has taken over the world from China and is shown as the biggest competitor of ChatGPT, may take its already good power consumption to a much higher level, according to new news. According to reports, the updated R1 AI model of the artificial intelligence application has reached a level where it can run on a single GPU. With the launch of DeepSeek-R1-0528-Qwen3-8B, a smaller version of the new R1 model, DeepSeek claims that this model has left similarly sized models behind in certain benchmark tests.

Based on the Qwen3-8B model that Alibaba launched in May, the smaller and updated R1 version performed much better than Google's Gemini 2.5 Flash model in the AIME 2025 test, which consists of challenging math questions.

DeepSeek-R1-0528-Qwen3-8B achieved almost the same performance as Microsoft's recently released Phi 4 model on another math skill test, HMMT.

So-called scaled-down models like DeepSeek-R1-0528-Qwen3-8B are generally less capable than their full-size counterparts, but they still stand out with their performance. The most important advantage of these versions, of course, is that they require much fewer computational resources.

According to cloud platform NodeShift, the Qwen3-8B requires a GPU with 40GB to 80GB of RAM (for example, an Nvidia H100) to run, while the full-size new R1 requires about a dozen 80GB GPUs.

DeepSeek-R1-0528-Qwen3-8B is available under a permissive MIT license, meaning it can be used commercially without restrictions. Many distributors, including LM Studio, offer this model via an API.