DeepSeek V3.2

• - 🆕 Open-weight reasoning model debuting DeepSeek Sparse Attention for long contexts.
• - 📏 Long-context window with near-linear attention cost via DSA.
• - 🧠 Strong math/coding reasoning, reported gold-level IMO and IOI results.
• - 🔧 Built for agentic tool use and function calling.
• - ⚡ DSA cuts attention complexity while preserving output quality.
• - 📚 Continued-trained from V3.1-Terminus, adding only the sparse-attention change.
• - 🔒 Released under the permissive MIT license.
• - 🌐 Supports reasoning, function-calling, and web-search capabilities.

DeepSeek

DeepSeek is a Chinese artificial intelligence company specializing in large language model development, founded in July 2023 by Liang Wenfeng. Based in Hangzhou, Zhejiang, the company is backed by High-Flyer, a prominent Chinese hedge fund also co-founded by…

DeepSeek V3.2 is an open-weight, reasoning-first large language model from Hangzhou-based DeepSeek, released in December 2025 under an MIT license. Its headline change is DeepSeek Sparse Attention (DSA), an efficient attention mechanism that substantially reduces computational complexity while preserving performance in long-context scenarios. According to the technical report, DSA uses a "lightning indexer" followed by fine-grained token selection, bringing attention cost toward near-linear scaling instead of the quadratic cost of traditional transformers.

Within the V3 family, V3.2 is the production successor to the experimental V3.2-Exp, which itself built on V3.1-Terminus by introducing DSA. According to DeepSeek's technical report, the only architectural modification versus V3.1-Terminus is the addition of sparse attention through continued training, and the new base model achieves performance on par with the previous iteration despite the efficiency change. On the independent Fiction.liveBench long-context evaluation, the report notes V3.2-Exp does not regress relative to V3.1-Terminus.

Beyond efficiency, DeepSeek's technical report pairs V3.2 with a scalable reinforcement-learning post-training framework and large-scale agentic task synthesis spanning many tool-use environments, aimed at stronger reasoning and function-calling. DeepSeek reports gold-medal-level results on 2025 competition benchmarks including the IMO and IOI.

V3.2 is the newest entry in this lineage before DeepSeek's later models DeepSeek V4 Pro and DeepSeek V4 Flash, both dated 2026, which represent the next generation beyond the V3 series.