DeepSeek’s New Model Get Pretty Good Marks On Thinking

There’s a new model in town, and it’s turning heads inside of the industry, although most Americans probably have not heard of it.

DeepSeek is a large language model developed by a firm called High Flyer Capital Management, which experts describe as a quantitative hedge fund. This organization has created a number of DeepSeek models and versions, some of which are out-performing the most current models under development at OpenAI and elsewhere.

In particular, there’s a new model called DeepSeek R1-Lite-Preview, which is getting top billing from some analysts, as one of the most effective tools out there right now in the beginning of our chain of thought reasoning era.

This DeepSeek model displays chain of thought, and does better on some types of logical inference, mathematical reasoning, and real time problem-solving tasks then what OpenAI or Anthropic can provide.

Domestic Competition and International Competition

Domestically, DeepSeek is contending with the Qwen models developed by Alibaba. I reported yesterday on how Qwen does fairly well with the MATH data set that uses high school math curriculum to test models and compare and contrast their abilities.

DeepSeek-R1-Lite-Preview is specifically credited with outperforming other models of its kind with the MATH data set and other prompting.

Internationally, OpenAI and Anthropic and other firms are trying to maintain dominance in the field. However, we have people like the CEO of Hugging Face, Clement Delangue, conceding that “(Qwen’s model) is the king, and Chinese open models are dominating overall.”

And then you have figures like Liu Qingfeng, founder of Chinese AI group iFlytek, who has reportedly said, “The gap between the US and China isn’t as big as everyone thinks. In a lot of verticals our [models] are better than theirs.”

There’s also the debate about open models that’s taking up a lot of time on the tech media stage. Closed models are safer, but open models are more egalitarian.

What DeepSeek’s Model Can Do

If you’re into the more intuitive details of how models outperform each other, you may be wondering: what does DeepSeek-R1-Lite-Preview do that’s so much better than other models?

This article at Venture Beat provides two very concrete examples of the model solving the question: “how many letter R’s are in the word strawberry?” and, alternately in numerical language, “which is larger, 9.11 or 9.9?”

So that’s the kind of question-answering that is showcasing problem-solving and logical inference abilities.

Specifically, Venture Beat reports that the model exceeds OpenAI-o1-Preview level performance in MATH.

Its reasoning capabilities are enhanced by its transparent thought process, allowing users to follow along as the model tackles complex challenges step-by-step.

To be fair, OpenAI’s o1 preview also does chain of thought. I’ve used it and been impressed by looking at the tasks that it’s doing on its way to answering a question. But it’s important to understand the market context, and which models are front and center right now, as the brightest LLMs come to the front of the classroom.

As for logical applications, we’re seeing any and all of these new models being integrated into industry tools in a wide spectrum of fields. But that’s a little bit about what analysts are saying right now, as they evaluate models that are still in beta, still very nascent, and still emerging into our world.