Chinese AI startup DeepSeek has introduced its newest model, DeepSeek-V3.2-Speciale, claiming it can perform some tasks as well as the latest models from Google DeepMind and OpenAI.

This is notable because DeepSeek has limited access to advanced computer chips used in AI training.

The new AI is designed to think and plan like an agent, meaning it can not only generate text, but also use tools such as search engines, coding programs, and other software to solve problems.

The standard version, DeepSeek-V3.2, is said to reach performance similar to “GPT-5 level” on general tasks. The special version, V3.2-Speciale, reportedly matches the reasoning skills of DeepMind’s Gemini 3 Pro in tasks that require logic and problem-solving.

DeepSeek also claims its model performs well in math and coding tests, including international competitions, and can handle tasks that require complex reasoning.

The model uses a new technology called Sparse Attention, which helps it process long inputs efficiently without needing extremely powerful hardware.

Experts say DeepSeek’s new model shows how Chinese AI startups are starting to challenge Western companies.

If the model performs as claimed, it could mark an important step toward “agentic AI”, where systems can not only generate information but also plan and act on it using external tools.

It remains to be seen how the model performs in independent tests, whether developers start using it worldwide, and how other AI companies respond.

Copied