DeepSeek's new V3.2 and Speciale, reasoning and performance rivaling GPT-5 and Gemini 3.

DeepSeek launched two models on December 2nd, namely DeepSeek-V3.2 and the DeepSeek-V3.2-Speciale, which focuses on mathematical reasoning. The official announcement stated that these two models continue the original experimental direction, aiming for a comprehensive upgrade in reasoning ability, tool usage, and long-term thinking capabilities. DeepSeek also emphasized that the official version V3.2 has demonstrated performance in multiple reasoning tests that is comparable to GPT-5 and Gemini-3 Pro, while the Speciale version has achieved international gold medal standards in mathematics and information competitions.

The experimental version leads the way, with the official version V3.2 taking over.

DeepSeek positioned its V3.2-Exp launched in September as an experimental platform towards the next generation of AI. The official version released this time, DeepSeek-V3.2, no longer carries the “Exp” designation, symbolizing a more mature functionality.

According to the official statement, the new version V3.2 performs similarly to GPT-5 and Gemini-3 Pro in multiple reasoning tests, and it is particularly emphasized that this is their first model that closely integrates “thinking modes” and “tool applications,” while supporting both thinking and non-thinking modes simultaneously. From the image below, it can be seen:

“DeepSeek-V3.2 has demonstrated its tool usage capabilities through benchmark test reports, placing it alongside top models such as GPT-5 and Gemini-3 Pro.”

The reasoning ability has been upgraded again, with tool integration being the biggest highlight.

DeepSeek indicates that a major highlight of V3.2 is the ability to combine the reasoning process with the use of tools. In other words, while the model is contemplating something, it can also call upon external tools such as search engines, calculators, and code executors, making the overall task handling process more complete, more autonomous, and closer to the way humans solve problems.

Speciale focuses on long reasoning, with mathematical performance reaching gold medal level.

In addition to the standard version V3.2, DeepSeek also launched another version DeepSeek-V3.2-Speciale. This version is specifically designed for high-difficulty mathematical reasoning and extended thinking.

The official positioning aims to explore the limits of open-source model inference capabilities and even see what boundaries the model itself can reach. In terms of results, Speciale achieved gold medal level in competitions such as the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI), with inference performance comparable to Google’s latest Gemini-3 Pro. As can be seen from the figure below:

“DeepSeek-V3.2-Speciale's reasoning capabilities have reached the gold medal level of international mathematics and information competitions, outperforming or matching GPT-5, Gemini-3 Pro, and Kimi-K2 in multiple reasoning and programming competition benchmarks.”

New training method revealed, AI agent capability further enhanced.

Outside of the model, DeepSeek has also made a public announcement of a new research achievement, which is that they have developed a new method for training AI agents. These agents can interact with the external environment, analyze data, and make judgments on their own without the need for continuous human instructions.

DeepSeek emphasizes that this is the foundational technology they designed to enable AI to operate more efficiently and respond more quickly.

Continuing the momentum of January, the pace of research and development continues to accelerate.

DeepSeek garnered global attention in January this year due to a groundbreaking model. The V3.2 series, released now, is their latest achievement following that success. Just before the launch of V3.2, DeepSeek released DeepSeekMath-V2 last week, an open-source model focused on mathematical theorem proofs, demonstrating their ongoing enhancement in reasoning and mathematics.

The technical report is out, and the inference capability of V3.2 is close to that of GPT-5 and Kimi.

DeepSeek also released a technical report titled “DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models,” indicating that V3.2 performs similarly to GPT-5 and Kimi-k2-thinking in several reasoning benchmark tests.

The report also emphasizes that the competitiveness of domestic open-source models in China in the field of inference remains on par with the world's top models.

This article introduces DeepSeek's new V3.2 and Speciale, which approach the reasoning and performance of GPT-5 and Gemini 3. It first appeared in Chain News ABMedia.