Axion 1.0: A 35B Open Agent Foundation Model with Simplified Training

Axion 1.0 is a new 35B open agent foundation model trained using a streamlined three-stage recipe that combines supervised fine-tuning and reinforcement learning. The developers released not only the model weights but also the complete training code, prompts, and data construction pipelines. This release demonstrates that powerful agentic capabilities can be achieved without complex training interventions like direct preference optimization or on-policy distillation. By fully open-sourcing the development pipeline, it provides a highly reproducible path for the open-source community to build and study advanced agentic models. Axion 1.0 avoids dedicated mid-training, DPO, and OPD, relying instead on two lightweight SFT stages focused on self-contained and tool-dependent reasoning, followed by reinforcement learning. It achieves performance competitive with leading commercial models across multiple agentic benchmarks.