Axion 1.0 is an open agent foundation model trained via a simplified three-stage recipe: two lightweight supervised fine-tuning stages and reinforcement learning. This method avoids expensive mid-training or DPO by using model-aware data construction and scalable tool-dependent reasoning pipelines.
HOW THIS AFFECTS YOU
●
builderYou can utilize a simplified training recipe to develop capable agentic models.
●
researcherThe three-stage training approach demonstrates effective agentic reasoning without DPO.