HybridAgent: the model thinks internally and uses 35–55% fewer tokens

The authors of arXiv paper 2610.18861 report an average quality gain of 10.4% across 11 benchmarks and 35.5–54.7% fewer tokens. The weights are already on Hugging Face: HybridAgent-8B and HybridAgent-14B, with roughly 980 downloads each as of Oct 13, 2026.