DeepSeek introduces open-source V3.2 models with Speciale variant… — Ars Dev — TG.ME

DeepSeek introduces open-source V3.2 models with Speciale variant matching Gemini-3.0-Pro on hard reasoning DeepSeek’s new V3.2 models arrive like a sequel that actually fixes the plot. The setup is simple: developers want open models that think, plan, and act with the precision of top proprietary systems. The problem is that long-context reasoning and agent workflows usually break when attention costs spike or post-training budgets run thin. The insight came from studying where open models fall short: slow attention, weak RL signals, and limited agent data. DeepSeek answers with a redesigned attention layer and a scaled reinforcement learning pipeline that treats reasoning as a first-class target. The standout moment comes from V3.2-Speciale, which reaches gold-level scores on the 2025 IMO, CMO, ICPC, and IOI and matches Gemini-3.0-Pro on complex reasoning. Key features and results: • DeepSeek Sparse Attention reduces long-context compute without hurting accuracy. • Reinforcement learning uses over 10% of pre-training compute to sharpen reasoning. • Agent data spans 1,800 environments and 85,000 prompts for stronger generalization. • V3.2-Speciale matches Gemini-3.0-Pro on demanding reasoning benchmarks. • Open weights on Hugging Face support fine-tuning with LoRA or full training.