DeepSeek V3.2
DeepSeek's late-2025 open-weight flagship, first to integrate thinking into tool-use
Shipped by DeepSeek · · class: open-weight-llm
DeepSeek released V3.2 on December 1, 2025, alongside a V3.2-Speciale reasoning variant. V3.2 is positioned as the general-purpose / speed model; Speciale handles deep logic, complex coding, and multilingual reasoning workloads.
What was new at release: V3.2 was the first open-weight model to integrate “thinking” directly into tool-use, and supports tool-use in both thinking and non-thinking modes. The MIT-licensed weights kept DeepSeek’s open-weight pressure on closed-frontier labs.
What’s been observed since: V3.2 has become a common substrate for self-hosted enterprise deployments where cost-per-token matters more than the last 5% of capability. DeepSeek’s April 2026 V4 preview (V4-Pro / V4-Flash) reset expectations, but V3.2 remains the deployed version most enterprises are still standardized on as of May 2026.