DeepSeek Releases V3.2 & V3.2-Speciale
Narrative
Reasoning-first models for agents. V3.2: Official successor to V3.2-Exp with thinking integrated into tool-use (thinking/non-thinking modes), massive agent data synthesis (1,800+ environments, 85k+ instructions). V3.2-Speciale: Maxed-out reasoning variant rivaling Gemini-3.0-Pro. Gold-medal performance on IMO, CMO, ICPC World Finals, IOI 2025.
Reality
Launched December 1, 2025. V3.2 immediately available on web, app, API (balanced speed/reasoning, GPT-5 level claimed). V3.2-Speciale API-only (temporary endpoint until Dec 15, 2025; no tool-use, higher token usage for evaluation/research). Tech report details thinking-in-tool-use breakthrough. Community adoption rapid; positioned as agent-ready daily driver.
Implication
Pushed open-source reasoning/agent capabilities forward at low cost. Demonstrated thinking/tool-use integration without proprietary data. Speciale's competition-level wins (e.g., IMO gold) reinforced Chinese labs' frontier parity. Trade-offs (token efficiency, temporary access) highlighted scaling challenges. Built hype for V4 coding pivot.