DeepSeek releases improved V3 AI model three months after initial launch

DeepSeek, a Chinese AI company, has released DeepSeek-V3-0324, an updated version of its V3 AI large-language model (LLM), which was originally launched in December 2024. The V3 model surprised the world due to its much lower training hardware requirements, shorter training time, and lower API costs, yet high performance versus competing AI such as OpenAI's GPT models.

The updated V3 model is a non-reasoning AI model, meaning it attempts to answer questions quickly and does not take extra time to think through tough problems, unlike DeepSeek's R1 model. Its 685-billion parameter size places it among the largest publicly available LLMs. The latest model is available for use under the MIT License.

DeepSeek-V3-0324 scored 5.3% to 19.8% higher on AI benchmarks than the original V3. Its performance is comparable to other top AI, such as GPT-4.5 and Claude Sonnet 3.7.

The updated model also has improved capabilities. One area of improvement is its refined ability to create appealing web pages and front-ends for online games. Another improvement is its ability to better search, write, and translate in Chinese.

Readers wanting to experiment with the full 685B DeepSeek-V3-0324 will need at least 700 GB of free storage space and multiple Nvidia A100/H100 GPUs. Distilled and smaller-parameter versions of the model can be run on a single GPU, such as the Nvidia 3090 (sold here on Amazon).