Researchers release Axion 1.0 open agent model

Researchers introduce Axion 1.0, a fully open agent foundation model trained through a simplified three-stage recipe consisting of two supervised fine-tuning stages and reinforcement learning. The approach avoids dedicated mid-training, direct preference optimization, or on-policy distillation while achieving competitive performance with leading commercial models across agentic benchmarks. The developers are releasing the complete training recipe and data construction pipeline to enable full reproducibility. Open article →