New Delhi: DeepSeek has launched its latest AI models, V3.2 and V3.2-Speciale, as the company steps up efforts to compete in the rapidly evolving global large language model landscape. The new releases position DeepSeek more firmly against leading systems such as GPT-5, Claude Sonnet 4.5 and Gemini 3 Pro, particularly in coding, reasoning and tool-use tasks.
The V3.2-Speciale model delivered notable performance in benchmark testing, achieving gold-level scores in the 2025 International Math Olympiad and Informatics Olympiad evaluations. This indicates continued progress in tackling advanced analytical and problem-solving tasks, an area that has become a key differentiator in the LLM race.
DeepSeek attributes the models’ efficiency gains to three technical updates: its DeepSeek Sparse Attention (DSA) mechanism, a scalable reinforcement learning framework and an expanded agentic task-synthesis pipeline. The company says DSA reduces computation demands for long-context processing while maintaining accuracy. Both models continue to use DeepSeek’s Mixture-of-Experts transformer architecture, featuring 671 billion total parameters with 37 billion active per token.
The update also refines the chat template and tool-calling protocols to support more structured reasoning workflows. With V3.2, DeepSeek aims to bolster its standing as an open-access alternative in an ecosystem increasingly dominated by proprietary, high-cost AI models.
