DeepSeek V3.2: A Low-Cost Contender for GPT-5

DeepSeek, a Chinese lab, has announced the release of V3.2, an open-source artificial intelligence model that promises to compete with GPT-5 in math reasoning tasks. The key feature of this model is its cost: it is estimated that running DeepSeek V3.2 costs about one-tenth compared to GPT-5.

This result was achieved through the implementation of a new architecture called "Sparse Attention." This architecture has made it possible to achieve high performance with an investment of approximately $5.5 million for training. In comparison, large US technology companies spend over $100 million to develop similar models.

DeepSeek's open-source approach could democratize access to advanced artificial intelligence models, allowing a wider audience to develop and deploy AI-based solutions. Competition in the field of large language models (LLMs) is growing rapidly, with new architectures and training techniques constantly emerging. This competition leads to faster progress and greater accessibility for users.