DeepSeek is back! Chinese startup launches two new models to take on Gemini and ChatGPT
China's DeepSeek launched V3.2 and V3.2-Speciale AI models, asserting performance comparable to GPT-5 and Google's Gemini 3 Pro. The new model achieved top scores in international math competitions, a feat that had been reserved for proprietary models so far.
China-based DeepSeek has launched two new AI models to take on the dominance of Gemini and ChatGPT in the chatbot space. Notably, the AI startup gained popularity at the start of the year with its DeepSeek R1 and V3 models that shattered all notions of Western supremacy in the AI race, since the models delivered similar performance to the likes of ChatGPT while being developed at a fraction of the cost.
Since then, however, DeepSeek hasn't been able to deliver similarly powerful models while Google's Gemini series and Nano series have dominated in text, image and video-related tasks. Meanwhile, OpenAI has also upped the ante with a bunch of agentic AI releases and model upgrades like the recently launched GPT 5.1. In contrast, other Chinese companies such as Qwen and Kimi have taken centre stage in the AI race with their powerful image and text generation models.
All of this might be about to change with the new DeepSeek V3.2 and DeepSeek V3.2 Speciale models.
What's new with DeepSeek V3.2 models?
DeepSeek says that the V3.2 series are reasoning- first models built for agents. The company says that DeepSeek V3.2 delivers GPT 5-level performance in general tasks.
The Chinese startup also claims that its DeepSeek V3.2 Speciale delivers reasoning abilities on par with Google's latest Gemini 3 Pro model, especially in complex problem-solving scenarios.
Moreover, the V3.2 Speciale model also managed to achieve gold medal performance in the 2025 International Mathematical Olympiad and International Olympiad in Informatics, a benchmark previously dominated only by proprietary models from Google and OpenAI.
DeepSeek V3.2 Speciale also outperforms both GPT 5 High and Gemini 3 Pro on pure math benchmarks like AIME and HMMT. The model also takes a lead in pure coding logic, showcased via CodeForces, where it defeats all other AI models on the list.
On Humanity's Last Exam, one of the hardest benchmarks going around, DeepSeek V3.2 Speciale scored higher than GPT 5 High but fell behind Gemini 3 Pro.
How to use the new DeepSeek models?
DeepSeek V3.2 has replaced the V3.2 Experimental model released in September as the default model on the chatbot's website and app. So, if you want to try out the new model, you just need to log in to the DeepSeek website or app and start using it.
Meanwhile, DeepSeek V3.2 Speciale is available only for API use for now. There is no update on when the model will make its way to regular users.
