DeepSeek: DeepSeek V3.2

DeepSeek-V3.2 is a cutting-edge large language model engineered for optimal computational efficiency and superior performance in reasoning and agentic tool-use. It integrates DeepSeek Sparse Attention (DSA), a novel fine-grained sparse attention mechanism that significantly reduces training and inference costs while maintaining high quality, especially in scenarios requiring long context windows. This model has demonstrated remarkable capabilities, achieving gold-medal results in the 2025 IMO and IOI, positioning its reasoning abilities in the GPT-5 class. Further enhancing its utility, V3.2 incorporates a scalable reinforcement learning post-training framework to refine reasoning and a large-scale agentic task synthesis pipeline. This pipeline improves the integration of reasoning into tool-use settings, boosting compliance and generalization in interactive environments. Users can fine-tune reasoning behavior using the `reasoning` `enabled` boolean. With a 163K token context window and 4K token max output, DeepSeek V3.2 is priced at $0.25/$0.38 per 1M input/output tokens, offering excellent value for its advanced capabilities.