The DeepSeek team released V3.2 ahead of the US holiday weekend, positioning it as a strong alternative to proprietary models like GPT-5 and Gemini 3.0 Pro.
V3.2 utilizes a non-standard sparse attention variant that requires custom code for inference, similar to the approach used in V3.1.
The release follows a shift from dedicated reasoning models (like R1) toward hybrid architectures, aiming to optimize performance across different use cases.
Source: https://magazine.sebastianraschka.com/p/technical-deepseek
