DeepSeek is great again馃憤! DeepSeekV3.2 is an open source large model designed for agent inference. Just announced open source! Through analysis, the DeepSeek team found that the open source model faces three major bottlenecks in complex tasks:
1. Reliance on traditional attention mechanisms leads to low efficiency in long text processing;
2. Insufficient computing resources in the fine-tuning stage, limiting the performance of difficult tasks;
3. Weak generalization ability and instruction following ability, affecting practical applications. To solve these problems, the DeepSeek team launched three key improvements: Optimize the architecture and introduce an efficient DSA attention mechanism to reduce computational complexity while ensuring long text performance; Develop a stable reinforcement learning protocol that allows the fine-tuning phase to significantly expand computing resources and release advanced capabilities; Propose a new tool reasoning pipeline to unify reasoning and tool usage tasks and introduce a large-scale agent task synthesizer to improve generalization and instruction following capabilities. Based on these innovations, DeepSeek-V3.2 has performed well in multiple benchmark tests and is close to the performance of closed-source systems, especially showing cost-effective advantages in agent tasks. The subsequently developed DeepSeek-V3.2-Special achieved performance parity with top closed-source systems in the field of inference and won gold medals in multiple international competitions. By properly optimizing the architecture, resource management and reasoning capabilities, the open source model can narrow the gap with the closed source model and provide a more economical solution.