China's DeepSeek Unveils New AI Models to Rival OpenAI and Gemini
DeepSeek has rolled out two advanced models it says can match OpenAI and Google on reasoning tasks, signalling another push from China’s AI sector to close the gap with Silicon Valley.
Topics
News
- Anthropic's Mythos Finds Critical Flaws in Classified US Systems: Report
- Qatar Launches AI-Focused Scholarship Program to Build Future Digital Workforce
- China's LineShine Tops TOP500, Becomes World's Fastest Supercomputer
- Japan's Sakana AI Unveils Fugu, Claims Edge Over Claude Fable 5 in Coding
- Oracle Cuts 21,000 Jobs as AI Silently Reshapes Its Operating Model
- Nvidia’s New Cooling System Cuts Data Center Water Use to Near Zero—But Not AI’s
Chinese AI startup DeepSeek has rolled out two AI models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, claiming performance on par with leading Western systems on complex reasoning and tool-use tasks.
The Hangzhou-based firm said DeepSeek-V3.2, its new flagship model, is the official successor to the experimental version showcased in September.
In internal tests, the company said the model matched OpenAI’s GPT-5 across several reasoning benchmarks, a notable assertion at a time when China’s open-source AI ecosystem is striving to keep pace with proprietary Western systems.
But the real spotlight is on DeepSeek-V3.2-Speciale, a high-compute, specialized variant built to push reasoning capabilities further.
According to DeepSeek, Speciale not only surpasses GPT-5 on reasoning tasks but also performs at a level comparable to Google’s Gemini-3.0-Pro.
The company said the model has demonstrated “gold-medal performance” in both the 2025 International Mathematical Olympiad and the International Olympiad in Informatics, two of the world’s toughest tests of mathematical and problem-solving ability.
DeepSeek credits its leap in performance to a mix of architectural and methodological advances: a “sparse attention” mechanism to reduce compute costs while preserving context depth; efficient “mixture-of-experts” routing that activates only part of the network per token; and a large-scale pipeline to synthesize agent-training data spanning thousands of simulated environments to improve generalization on complex tasks.
DeepSeek-V3.2 is already available via the company’s app, web interface and APIs. The Speciale variant is for now accessible only through a limited-access API, signaling that DeepSeek is positioning it as a research-grade or high-intensity computational tool rather than a mass-market product.
