DeepSeek Debuts V3.2 and V3.2-Speciale, AI Models Optimized for Agent-Based Reasoning

December 1, 2025

Details

  • On December 1, 2025, DeepSeek introduced two advanced AI models: DeepSeek-V3.2 and the high-performance DeepSeek-V3.2-Speciale.
  • V3.2 serves as the successor to V3.2-Exp and is immediately accessible on DeepSeek’s web, mobile, and API platforms at no additional cost.
  • The V3.2-Speciale is available as a limited API-only preview through a temporary URL until December 15, 2025, with the same pricing as V3.2 but without tool call functionality during the preview period.
  • Benchmark data suggests V3.2 achieves “GPT-5-level” performance, while Speciale competes with Google’s Gemini-3.0-Pro, both securing top ranks on high-profile reasoning test suites like IMO, CMO, and ICPC World.
  • Both models leverage a novel training dataset that spans over 1,800 environments and 85,000 intricate instructions, enhancing their capabilities in multi-step reasoning and API-driven task planning.
  • DeepSeek has open-sourced the model weights, tokenizer, and technical documentation, inviting community-driven fine-tuning and independent assessments.

Impact

DeepSeek’s strong evaluation scores put pressure on leading competitors like OpenAI and Google as they prepare upcoming model launches. By making its models freely available, DeepSeek could accelerate innovation in autonomous agent development, drawing interest from startups and researchers. The open release also addresses anticipated regulatory requirements and could influence the direction of future AI research and policy.

Read more