DeepSeek-V3.2 Outperforms GPT-5 on Reasoning Tasks

"DeepSeek applied three new techniques in the development of DeepSeek-V3.2. First, they used a more efficient attention mechanism called DeepSeek Sparse Attention (DSA) that reduces the computational complexity of the model. They also scaled the reinforcement learning phase, which consumed more compute budget than did pre-training. Finally, they developed an agentic task synthesis pipeline to improve the models' tool use."

"First, due to fewer total training FLOPs, the breadth of world knowledge in DeepSeek-V3.2 still lags behind that of leading proprietary models. We plan to address this knowledge gap in future iterations by scaling up the pre-training compute. Second, token efficiency remains a challenge...Future work will focus on optimizing the intelligence density of the model's reasoning chains to improve efficiency."

DeepSeek-V3.2 is a family of open-source reasoning and agentic AI models whose high-compute variant, DeepSeek-V3.2-Speciale, outperforms GPT-5 and matches Gemini-3.0-Pro on several reasoning benchmarks. Three principal techniques were introduced: DeepSeek Sparse Attention (DSA) to lower computational complexity, a scaled reinforcement learning phase that used more compute than pre-training, and an agentic task synthesis pipeline to improve tool use. The model leads most open models across coding, reasoning, and agentic benchmarks but retains gaps in world knowledge, token efficiency, and complex-task performance compared with frontier closed-source systems.

Read at InfoQ