Chinese AI startup DeepSeek has openly released two new large language models that rival the capabilities of leading systems from OpenAI and Google while being dramatically more efficient and completely free to use. The models, DeepSeek-V3.2 and a specialized competition variant called V3.2-Speciale, each feature 685 billion parameters.
According to the company's technical report, the models achieve performance that "matches or surpasses" GPT-5 and Gemini 3.0 Pro on elite academic benchmarks. The specialized variant reportedly won a gold medal at the 2025 International Mathematical Olympiad (IMO) and placed second in the International Collegiate Programming Contest (ICPC) World Finals. The base model is positioned as a highly capable coding and reasoning assistant.
A key innovation is the DeepSeek Sparse Attention (DSA) mechanism, which the company states reduces the computational cost of processing long documents by up to 70% compared to its predecessor. This makes operating the massive model significantly cheaper.
In a major shift for frontier AI, the models are released under a permissive MIT license, allowing full public access to the model weights for any purpose, including commercial use. This open-source strategy directly challenges the closed, API-based business models of Western AI leaders and is set to intensify global AI competition. The release has already attracted regulatory scrutiny in Europe, with Germany raising data governance concerns.