DeepSeek-V3.2 is now in release, integrating thinking directly into tool-use for the first time, improving its predecessor DeepSeek-V3.2 Experimental. The model supports tool-use in both thinking and non-thinking modes. China-based DeepSeek began disrupting the U.S. AI market in January with the debut of foundation models that rival those from Google and OpenAI that are available for free. The company released internal benchmark scores indicating its new model can compete with OpenAI’s GPT-5 in reasoning benchmarks and agentic tasks. A variation, DeepSeek-V3.2-Speciale, has been released for specialized math and is said to perform comparably to Google’s Gemini 3 Pro.

DeepSeek says in an X post that V3.2-Speciale “dominates complex tasks but requires higher token usage,” and is currently available via paid API only “to support community evaluation and research,” while the basic model can be accessed for free on the web and through the DeepSeek app.

“DeepSeek also said the V3.2 version combines that ability to mimic something like human reasoning with the capability to use tools like search engines, calculators and code executors,” writes Bloomberg, adding that “the second version DeepSeek released, V3.2-Speciale, focuses on mathematical computations and other long-thinking capabilities.”

DeepSeek said its goal is “to ‘push the inference capabilities of open-source models to their limits and explore the boundaries of model capabilities,” Bloomberg reports.

The company writes on its Hugging Face page that V3.2’s “high computational efficiency with superior reasoning and agent performance” is built on three key technical breakthroughs:

1. DeepSeek Sparse Attention (DSA): an efficient attention mechanism that substantially reduces computational complexity while preserving model performance, specifically optimized for long-context scenarios, allowing V3.2-Speciale to surpass GPT-5.
2. Scalable Reinforcement Learning Framework: implementing a robust RL protocol and scaling post-training compute, that allows DeepSeek-V3.2 to perform comparably to GPT-5.
3. Large-Scale Agentic Task Synthesis Pipeline: integrating reasoning into tool-use scenarios, using a novel synthesis pipeline that systematically generates training data at scale, facilitating scalable agentic post-training to improve compliance and generalization in complex interactive environments.

VentureBeat reports DeepSeek designed V-3.2 “as an everyday reasoning assistant,” and V3.2-Speciale as a high-powered counterpart “that achieved gold-medal performance in four elite international competitions: the 2025 International Mathematical Olympiad, the International Olympiad in Informatics, the ICPC World Finals, and the China Mathematical Olympiad.”

If independently verified, the performance of these new models as demonstrated in DeepSeek testing “could reshape the competitive landscape between American tech giants and their Chinese challengers,” VentureBeat suggests.