Axion 1.0: Open Agent Foundation Models, Simplified

Axion 1.0. Open Agent Foundation Models, Simplified. By Yuqi Jia and six others, from Duke University, ByteDance, and two other institutions. Published on the second of November.

Building an agent foundation model, one with strong and broad agentic capabilities, often involves a long chain of training stages. The abstract names the pieces Axion leaves out: dedicated mid-training, direct preference optimization, and on-policy distillation. Each extra stage adds complexity, and a long, partly hidden recipe is also hard for anyone else to reproduce.

Axion 1.0 is a fully open agent foundation model built on a remarkably simple recipe. It is trained in only three stages: two lightweight supervised fine-tuning stages, followed by reinforcement learning. The claim is that this compact path is enough to reach strong agentic performance, and that the whole of it can be made public. Each stage has a distinct job. The first teaches the model to reason on its own. The second teaches it to reason with tools. The third, reinforcement learning, stretches both across longer horizons and real interaction. There is no separate stage for preference data and no teacher policy to distil from.

The first fine-tuning stage builds strong self-contained reasoning. Its data is model-aware. The authors synthesize challenging problems, then adversarially strengthen them based on the base model's own successful solution strategies, so the problems are shaped around what the model can already do. The data, in other words, is built for this particular base model rather than collected generically, and it is made harder in exactly the places where that model already finds its way to an answer.

The second stage extends those capabilities to reasoning that depends on tools. It uses a scalable pipeline for constructing diverse tasks that can be verified automatically, grounded in real-world data sources and in deterministic computation. Reinforcement learning then comes last, further strengthening long-horizon reasoning and interaction with external tools and environments.

Despite the compact recipe, Axion 1.0 achieves strong performance across a broad suite of agentic benchmarks, and the abstract reports it is competitive with leading commercial models on several of them. Openness is the other half of the result. The release goes beyond model weights to the complete recipe: data construction, generation, filtering, verification, and prompts, along with the training pipeline, so that the entire development process can be reproduced and extended.

The argument Axion makes is about what can be removed. Three stages instead of many, no preference optimization, no distillation from another policy, and a recipe published in full. If strong agent models can be both open and that simple, the authors suggest, there is a practical and reproducible path forward for open agentic intelligence, one that others can follow step by step rather than guess at.