AI lab DeepSeek has unveiled two new reasoning-first models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, expanding its suite of agent-based, tool-integrated, and complex reasoning systems.
Both models and their accompanying technical reports have been released as open-source on Hugging Face.
The company announced on X that V3.2 serves as the official successor to V3.2-Exp and is now available across its mobile app, web interface, and API. The Speciale variant is accessible exclusively via a temporary API endpoint until December 15, 2025.
DeepSeek states that V3.2 is engineered to strike a balance between reasoning efficiency and long-context performance, describing it as “your daily driver with GPT-5-level capabilities.”
The V3.2-Speciale variant targets high-end reasoning tasks and is positioned to “compete directly with Gemini-3.0-Pro.” According to DeepSeek, Speciale achieves gold-tier (expert-human proficiency) results on competitive benchmarks such as the International Mathematical Olympiad (IMO), Chinese Mathematical Olympiad (CMO), and the ICPC World Finals.
These models build upon DeepSeek’s agent training methodology, leveraging a newly synthesized dataset comprising over 1,800 environments and 85,000 complex instructions. DeepSeek notes that V3.2 is its first model to directly integrate chain-of-thought reasoning into tool usage, enabling structured reasoning both within and alongside external tools.
Concurrent with the release, DeepSeek updated its API, confirming that V3.2 maintains the same usage patterns as its predecessor. The Speciale variant is priced identically to V3.2 but does not support tool calling. The company also highlighted a new capability in V3.2 called “Reasoning-in-Tool-Use,” with further details available in the developer documentation.
Recently, the company also launched another open-weight model, DeepSeekMath-V2. According to the AI lab, this model demonstrates exceptional mathematical theorem-proving abilities and earned a gold-tier score in the 2025 International Mathematical Olympiad (IMO).
