Ankit Sharma

DeepSeek-V3.2 - Reasoning-first models built for agents

by
We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Add a comment

Replies

Best
Ankit Sharma
Hunter
📌

Hey Hunters, I am excited to hunt DeepSeek-V3.2 & V3.2-Speciale! 🔥

DeepSeek just dropped two powerful reasoning-first models built especially for agents — and they are seriously impressive.

🚀 What’s new?

  • DeepSeek-V3.2: The official successor to V3.2-Exp — your new daily-driver with GPT-5-level capability.

  • DeepSeek-V3.2-Speciale: A max-reasoning beast that rivals Gemini-3.0-Pro. API-only for now.

🧠 World-Class Reasoning

  • V3.2-Speciale hits gold-level performance across IMO, CMO, ICPC World Finals & IOI 2025.

  • Designed for deep reasoning, long-chain planning & complex multi-step tasks.

🤖 Thinking in Tool-Use

  • New agent-training pipeline with 1,800+ environments & 85k+ complex instructions.

  • DeepSeek-V3.2 integrates “thinking mode” directly into tool-use — powerful for builders.

💻 API Highlights

  • Same usage as V3.2-Exp.

  • Speciale served via a temporary endpoint until Dec 15, 2025.

  • Same pricing, but uses more tokens due to advanced reasoning.

🛠 Open-Source Release

Both models are fully available on Hugging Face for developers to tinker with.

If you’re building agents, automation, or reasoning-heavy systems — this update is a game-changer. Happy to answer any questions! 🚀🔥

Gary Sztajnman
Claude opus 4.5 needs to be in these benchmarks