DeepSeek: DeepSeek V3.2 Exp

DeepSeek-V3.2-Exp is an experimental large language model from DeepSeek, serving as an intermediate step towards future architectures. It introduces DeepSeek Sparse Attention (DSA), a novel fine-grained sparse attention mechanism. DSA is engineered to significantly enhance training and inference efficiency, particularly in long-context scenarios, without compromising output quality. Users can fine-tune reasoning behavior using the `reasoning` `enabled` boolean. This model is ideal for exploring efficient transformer designs and advanced AI research. Developed under conditions aligned with V3.1-Terminus, DeepSeek-V3.2-Exp allows for direct performance comparisons. Benchmarking indicates performance generally on par with V3.1 across critical domains such as reasoning, coding, and agentic tool-use, with minor variations. Its primary focus is on validating architectural optimizations for extended context lengths rather than raw task accuracy, making it a valuable tool for researchers. Key specifications include a substantial 163K token context window and a 4K token maximum output. It supports functions, code generation, and streaming, excelling in code, reasoning, and math tasks. Pricing is competitive at $0.21/$0.32 per 1M input/output tokens. Access is available at the STARTER tier.