Deepseek’s new AI is smarter, faster, cheaper, and a real rival to OpenAI's models

• Chinese AI startup DeepSeek has released an upgraded AI model called V3-0324 to Hugging Face
• V3-0324 offers improved reasoning and coding abilities over its predecessors
• DeepSeek claims its AI models can match or beat those of American AI developers like OpenAI and Anthropic

DeepSeek dropped a major upgrade to its AI model this week, which has people buzzing almost as much as they did when the Chinese AI startup first made its splash earlier this year. The new DeepSeek-V3-0324 model is now live on Hugging Face, setting up an even starker rivalry with OpenAI and other AI developers.

According to the company's tests, DeepSeek's new iteration of its V3 model boasts measurable boosts in reasoning and coding ability. Better thinking and coding might not sound revolutionary on their own, but the pace of improvement and DeepSeek's plans make this release notable.

Formed just last year, DeepSeek has been moving fast, starting with the December release of the original V3 model. A month later, the R1 model for more comprehensive research debuted. Now comes V3-0324, named for its March 2024 release.

DeepSeek demand

The improvements bring the model to near-parity with OpenAI’s GPT-4 or Anthropic’s Claude 2 models. But, even if they aren't quite the same power, they run a lot cheaper, according to DeepSeek.

That's ultimately a huge selling point as AI use, and thus AI costs, continue to increase. Training AI models is notoriously expensive, and OpenAI and Google have huge cloud budgets that most companies couldn't reach without partnerships like OpenAI's with Microsoft. That exclusivity vanishes if DeepSeek's cheaper achievements become more common.

U.S. dominance of AI models is starting to slip anyway, thanks in part to Chinese startups like DeepSeek. It no longer seems shocking when the hottest model emerges from Shenzhen or Hangzhou. Geopolitical considerations, as well as business concerns, have spurred calls to ban DeepSeek from at least the U.S. government.

You probably won't see DeepSeek’s latest release changing everything for your schedule tomorrow, though. It hints that the ballooning demand for computational power and energy to fuel next-generation AI might not be as staggering as feared.

It also just might mean that the AI chatbot rewriting your resume or debugging your website also speaks fluent Mandarin.

You might also like

• I discovered a surprising difference between DeepSeek and ChatGPT search capabilities
• OpenAI calls on US government to ban DeepSeek, calling it ‘state-subsidized’ and ‘state-controlled’
• I pitted Gemini 2.0 Flash against DeepSeek R1, and you might be surprised by the winner