DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism that reduces training and inference cost while preserving quality in long-context scenarios. A scalable reinforcement learning post-training framework further improves reasoning, with reported performance in the GPT-5 class, and the model has demonstrated gold-medal results on the 2025 IMO and IOI. V3.2 also uses a large-scale agentic task synthesis pipeline to better integrate reasoning into tool-use settings, boosting compliance and generalization in interactive environments. Reasoning is on.
Evaluations
7
across 7 benchmarks
Latency
662ms
Context length
128k
tokens
Cost
$0.28 · $0.42
input · output per 1M tokens
Key takeaways
DeepSeek V3.2 (Reasoning) is designed for high computational efficiency, strong reasoning, and agentic tool-use, excelling in functional code generation and system administration tasks requiring complex logical deductions and interactive agentic capabilities.
It is a Transformer-based model incorporating DeepSeek Sparse Attention (DSA) for efficiency in long contexts, trained using a scalable reinforcement learning post-training framework and an agentic task synthesis pipeline.
The model demonstrates strong capabilities in understanding and executing explicit, well-defined technical instructions, coding (especially common Python programming functionality), and system administration, achieving 94.67% accuracy in code generation. It excels in tasks involving basic arithmetic, list/string manipulation, and common programming concepts.
Known limitations include struggles with complex multi-agent/interactive environments (e.g., maze exploration, Zork game, achieving a 42% passing score on Terminal-Bench), nuanced debugging, specialized search algorithms, and continuous state management.
The model performed exceptionally well on the AIME 2024 benchmark, achieving an Exact Match Accuracy of 93.33%, primarily demonstrating strengths in mathematical reasoning. It struggled significantly on AIME 2026 Part II problems, scoring 0%.
