DeepSeek-v3.2
10 days ago
- #AI
- #Machine Learning
- #DeepSeek-V3.2
- DeepSeek-V3.2 introduces three key technical breakthroughs: DeepSeek Sparse Attention (DSA), Scalable Reinforcement Learning Framework, and Large-Scale Agentic Task Synthesis Pipeline.
- DeepSeek-V3.2-Speciale variant surpasses GPT-5 and matches Gemini-3.0-Pro in reasoning.
- Achieved gold-medal performance in 2025 IMO and IOI.
- New chat template updates include revised tool calling and 'thinking with tools' capability.
- Includes Python scripts for encoding and parsing messages in OpenAI-compatible format.
- Local deployment recommended with temperature = 1.0, top_p = 0.95.
- DeepSeek-V3.2-Speciale is optimized for deep reasoning tasks and lacks tool-calling functionality.
- Licensed under MIT License.