Axion 1.0 details a simplified recipe for open agent models

Axion 1.0 is trained first with supervised fine-tuning, then with a second supervised fine-tuning stage, and finally with reinforcement learning.[1] The paper says the recipe does not use dedicated mid-training, direct preference optimization or on-policy distillation.[1] Rather than treating those omitted methods as a result in themselves, teams assessing the work can use them to understand the authors’ chosen training design. The paper describes the problems as synthesized and adversarially strengthened based on successful solution strategies from the base model.[1] That makes problem construction part of the recipe, not just a final evaluation step.

The second fine-tuning stage trains tool-dependent reasoning on diverse tasks that can be automatically verified, grounded in real-world data and deterministic computation.[1] The final reinforcement-learning stage is intended to strengthen long-horizon reasoning and interaction with tools and environments.[1] For teams evaluating agent models, these details indicate what capabilities the training process is designed to target; they do not, by themselves, establish performance in a particular deployment. The first-stage dataset contains 10,819 scientific and mathematical reasoning trajectories, while the second contains 7,544 trajectories and about 312 million tokens.[3]

The project website describes two main mixture-of-experts models, with 35 billion and 122 billion parameters.[2] The listed parameter figures are 35B / 3B and 122B / 10B, respectively.[3] The project also reports 4B and 9B variants trained only with supervised fine-tuning, and says its supervised fine-tuning runs use a 128K context window.[3] The authors say the development recipe release includes data construction, generation, filtering, verification, prompts and training code.[1] The Hugging Face page identifies the 35B checkpoint as a community reproduction, not an official Axion checkpoint, and lists that model under Apache License 2.0.[3][4]

The project reports a score of 59.0 for Axion 1.0-122B on full HLE with restricted website access.[3] It also reports 62.2 for Axion 1.0-35B on the SEAL-0 agentic problem-solving benchmark and 87.0 for Axion 1.0-122B on FrontierScience Olympiad.[3] These are reported benchmark results, not evidence here of performance in a specific workplace task. The project notes that some comparison-model scores were evaluated by the Axion team, while other baseline scores came from official reports.[3] Readers comparing results should therefore check how each score was produced.