DeepSeek’s V3 AI model gets a major upgrade – here’s what’s new

Not that it ever left, but it appears Chinese AI startup DeepSeek is back in the news — this time with an updated version of its V3 model, released in December.

Also: The best AI chatbots

On Tuesday, the company officially announced V3-0324, named after its release month and day. A day earlier, people noticed DeepSeek had uploaded the new model to HuggingFace, but with little additional information. 

What’s new in DeepSeek’s V3-0324 model?

Like R1 — DeepSeek’s top-performing model released in January and an OpenAI competitor — the new version is open source (in that its weights are public, not its actual code) under an MIT license.

In a post on X, DeepSeek noted that the update shows better coding skills for web development and a “major boost in reasoning performance,” but it still recommends it be used for less complex reasoning tasks. R1 remains the lab’s top reasoning model, ranking in fourth place on the Chatbot Arena.

Also: What is DeepSeek AI? Is it safe? Here’s everything you need to know

DeepSeek said the update shows improved performance over V3 on several industry-standard benchmarks, most notably the AIME (American Invitational Mathematics Examination) math benchmark, scoring nearly 20 points higher.

While benchmarks have become too easy for most models, a problem known as benchmark saturation, AIME is still considered more challenging than most. In January, Scale AI and the Center for AI Safety (CAIS) released Humanity’s Last Exam to combat saturation.

Also: This new AI benchmark measures how much models lie

That said, because it is based on high school math content, AIME’s answers are publicly available online, meaning they can be included in training data.

Also: Is OpenAI doomed? Open-source models may crush it, warns expert

According to DeepSeek, other improvements include “enhanced” writing style and improved quality, especially for longer-form content. Some Reddit commenters are speculating that the release of the upgrade could foreshadow the arrival of R2, which is anticipated to be as disruptive as R1.

How to try DeepSeek’s V3-0324 model

You can access V3-0324 now via HuggingFace or directly through DeepSeek’s website and app, though you may want to consider the major security holes and user privacy concerns first. While V3 and R1 proved to be very easily and dangerously jailbroken, it’s unclear as of now whether DeepSeek added any layers of security in V3-0324.

Also: How to try DeepSeek R1 – without the censorship or security risk

Want more stories about AI? Sign up for Innovation, our weekly newsletter.