DeepSeek-V3. 2: Pushing the Frontier of Open Large Language Models DeepSeek-V3 2 introduces DeepSeek Sparse Attention (DSA) for efficient long-context processing and a scalable RL framework that achieves performance comparable to or exceeding state-of-the-art models like GPT-5 and Gemini-3 0-Pro