DeepSeek Launches New AI Models Challenging OpenAI and Google

DeepSeek has released two new AI models that bring the business closer to the performance levels set by OpenAI and Google. According to the statistics that came out with the introduction, the new systems, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, had benchmark scores that were the same as or better than those of GPT-5 and Gemini 3 Pro.

The V3.2 model may now be used on all of DeepSeek's platforms and APIs. The Speciale version can only be used via API channels at first. The debut comes after the company's R1 model, which got a lot of attention earlier this year for being comparable with U.S. models.

DeepSeek claimed in a technical paper that came out with the release that open-source models have not been as good as closed-source systems for long-sequence efficiency, post-training computation, and following instructions in the real world. The business spoke about three major technological innovations that would help fill these gaps: a new sparse attention mechanism, a bigger reinforcement learning framework, and an agentic task-synthesis pipeline.

DeepSeek asserted that its high-compute Speciale variation could reason as well as the best proprietary AI models. It even won top spots in the 2025 International Mathematical Olympiad and the International Olympiad in Informatics. The findings show that worldwide AI developers are competing more and more as open-source projects reach performance levels that were only seen in major commercial laboratories before.