Model details
DeepSeek-V3.2 is engineered to harmonize high computational efficiency with advanced reasoning and agentic capabilities. At the core of its architecture is the DeepSeek Sparse Attention mechanism, a two-stage process that utilizes a high-speed lightning indexer to scan tokens and a fine-grained selection process to focus only on the most relevant information. This design significantly reduces computational complexity, allowing the model to maintain high performance during long-context scenarios while remaining efficient for daily use.
The model benefits from a robust reinforcement learning framework and a large-scale agentic task synthesis pipeline, which generates training data across thousands of environments to improve instruction-following. This post-training approach enables the model to integrate thinking directly into tool-use workflows. While the standard version serves as a balanced, high-performance daily driver, the high-compute Speciale variant pushes these capabilities further, achieving gold-medal results in competitive mathematics and informatics benchmarks.
Transparent token rates
Rates are shown per one million tokens. Combined means one million input plus one million output tokens.
SiliconFlow
DeepSeek says both models are more efficient and performant than DeepSeek V3.2 due to architectural improvements, and have almost "closed the gap" with current leading models, both open and closed, on reasoning benchmarks.
SiliconFlow
Cline announced that DeepSeek-V3.2 and the higher-reasoning V3.2-Speciale variant are available through its agentic coding tool, framing both as DeepSeek's first models explicitly designed for agentic workflows. The integration leverages the new thinking-in-tool-use behavior, where reasoning persists across multiple to The blog cites DeepSeek's agentic training scale: 1,800+ synthesized environments and 85,000+ complex agent instructions, broken into roughly 24,667 code-agent tasks, 50,275 search-agent tasks, 5,908 code-interpreter tasks in Jupyter, and 4,417 general agent tasks. Cline also notes that DeepSeek allocated over 10% of p
SiliconFlow
FriendliAI's model endpoint page documents deepseek-ai/DeepSeek-V3.2 as a 671B-parameter Mixture-of-Experts hybrid reasoning model released under the MIT license. It describes three core technical pillars: DeepSeek Sparse Attention for long-context efficiency, a scalable reinforcement learning post-training framework t The page also covers operational details relevant to API consumers: reasoning is disabled by default and toggled via a chat template flag, with reasoning traces returned through the reasoning_content field for multi-turn coherence. The chat template was updated to support revised tool calling and a new thinking-with-to
SiliconFlow
The official DeepSeek API documentation announces DeepSeek-V3.2-Exp, an experimental model released September 29, 2025, built on the V3.1-Terminus checkpoint and debuting DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism designed to speed up training and inference on long context. The page stat The release ships as fully open source: model weights are published on Hugging Face under the deepseek-ai organization, the tech report is hosted on GitHub, and key GPU kernels are provided in both TileLang (for rapid research prototyping) and CUDA. Together these artifacts establish the V3.2 family as a transparent, d
SiliconFlow
Compare DeepSeek-R1 and DeepSeek-V3.2 across performance, cost, capabilities, and real-world use cases. See which model fits your needs.
SiliconFlow
Compare DeepSeek-V3.1 and DeepSeek-V3.2-Exp across performance, cost, capabilities, and real-world use cases. See which model fits your needs.
SiliconFlow
Compare DeepSeek-V3 and DeepSeek-V3.2 across performance, cost, capabilities, and real-world use cases. See which model fits your needs.
SiliconFlow
DeepSeek released DeepSeek-V3.2, a family of open-source reasoning and agentic AI models. The high compute version, DeepSeek-V3.2-Speciale, performs better than GPT-5 and comparably to Gemini-3.0-Pro
