Axion 1.0: Open Agent Foundation Model Trained with Simple Three-Stage Recipe

Details

Axion 1.0 is a fully open agent foundation model released by Axion1-0, featuring strong agentic capabilities and a simple training recipe.
Training consists of three stages: two lightweight supervised fine-tuning (SFT) stages followed by reinforcement learning (RL), without dedicated mid-training, direct preference optimization (DPO), or on-policy distillation (OPD).
The first SFT stage develops self-contained reasoning through model-aware data construction, synthesizing challenging problems and adversarially strengthening them based on the base model's successful strategies.
The second SFT stage extends capabilities to tool-dependent reasoning using a scalable pipeline for diverse, automatically verifiable tasks grounded in real-world data and deterministic computation.
RL further strengthens long-horizon reasoning and interaction with external tools and environments.
Axion 1.0 achieves strong performance across a broad suite of agentic benchmarks and is competitive with leading commercial models on several of them.
The release includes the complete recipe: data construction, generation, filtering, verification, prompts, and training code, enabling full reproduction and extension.
Resources include an arXiv paper, GitHub repository, Hugging Face dataset, and project website.

Key point