Abstract:

The paper introduces DeepSeek-V3.2, an open large language model designed to improve computational efficiency and performance on reasoning and agentic tasks. The authors identify three main weaknesses in open-source models: inefficient attention mechanisms for long contexts, insufficient post-training compute, and poor generalization in agentic scenarios. Their solutions are: (1) a "DeepSeek Sparse Attention" (DSA) mechanism to reduce computational complexity from O(L²) to O(Lk), (2) a scalable reinforcement learning framework that dedicates over 10% of the pre-training cost to post-training, and (3) a pipeline for synthesizing a large volume of agentic tasks to improve tool-use. The paper claims DeepSeek-V3.2 performs comparably to models like GPT-5, while a high-compute variant, DeepSeek-V3.2-Speciale, allegedly achieves performance on par with Gemini-3.0-Pro, winning "gold medals" in top math and informatics competitions.

DeepSeek-V3.2: Key Technical Claims and Performance

• 1. The Problem: The authors claim the performance gap between closed-source and open-source LLMs is widening due to inefficiencies in attention mechanisms, insufficient post-training, and poor agentic capabilities in open models.
• 2. DeepSeek Sparse Attention (DSA): To improve efficiency, the model uses a sparse attention mechanism.
  • It employs a "lightning indexer" to calculate relevance scores and a "fine-grained token selection" mechanism to attend to only the top-k most relevant previous tokens.
  • This reduces the core attention complexity for long sequences, which they claim leads to significant cost savings in inference, particularly for long contexts.
• 3. Scalable Post-Training: The model's performance is heavily reliant on a massive reinforcement learning (RL) post-training phase.
  • They use Group Relative Policy Optimization (GRPO) and invest a computational budget exceeding 10% of the initial pre-training cost.
  • Stability is maintained through techniques like an unbiased KL estimate, off-policy sequence masking, and preserving expert routing paths and sampling masks between inference and training.
• 3.2 Agentic Task Synthesis: To improve tool-use, the model was trained on a large, diverse set of synthesized tasks.
  • They created over 1,800 distinct environments and 85,000 complex prompts for various agent tasks (search, code, general).
  • This is intended to improve the model's ability to generalize its reasoning to new, out-of-domain agent scenarios.
• 4. Performance Claims: The paper presents extensive benchmarks, though they are self-reported.
  • DeepSeek-V3.2: Claimed to be competitive with GPT-5-High and Kimi-K2-Thinking on various reasoning and agent benchmarks.
  • DeepSeek-V3.2-Speciale: A variant trained with fewer token constraints on reasoning data. It supposedly achieves performance on par with Gemini-3.0-Pro and gold-medal level results in the 2025 IMO, IOI, and ICPC World Finals. As far as I can tell, this comes at the cost of very poor token efficiency.
• 5. Limitations: The authors acknowledge several limitations.
  • The model has less world knowledge than top proprietary models due to a lower pre-training compute budget.
  • Token efficiency is a problem, especially for the "Speciale" version, which requires more tokens to achieve its results.
  • It is still considered inferior to frontier models on the most complex tasks.