This might be the biggest AI plot twist of 2025. Alibaba has just unveiled Tongyi DeepResearch, a 30B-parameter agentic model that outperforms GPT-4 and DeepSeek-V3 in deep research, while activating only 3.3B parameters at runtime. • Fully open-source. • Trained on just 2 H100s. The secret? A new agentic mid-training phase, where the model learns to reason, plan, and utilize tools before fine-tuning. Instead of predicting the next token, it learns to think like an agent. All synthetic data. 32K+ token samples. 10+ tool calls per task. At @spheron, we see this as the future, where compute efficiency, not compute dominance alone, defines the next era of AI. With decentralized, on-demand GPU access, breakthroughs like this can happen anywhere, not just inside trillion-dollar data centers. The future of AI is open, efficient, and powered by the people. Source: