Axion 1.0 simplifies Open Agent Foundation Models.
Axion 1.0 simplifies open agent models with 18,000 examples, lightweight stages, and reinforcement learning for scientific reasoning and tool use.
Key Takeaways
- Axion 1.0 simplifies open agent foundation-model training with about 18K supervised fine-tuning examples, two lightweight SFT stages, and reinforcement learning for scientific reasoning and tool use.
- The training recipe targets three goals: building challenging scientific problems for reasoning, using verifiable retrieval/tool-use tasks for investigation, and improving through RL over full agent trajectories.
- Axion 1.0-35B and 122B aim for competitive results across six evaluation settings: HLE, BioMysteryBench, SEAL-0, HiPhO, and the FrontierScience Olympiad and Research splits.
- The 35B reproduction checkpoint reports key scores such as 52.8 on HLE and 85.0 on FrontierScience Olympiad, while the release does not provide separate 122B benchmark scores in the same list.
- Apache 2.0 licensing, together with the paper, project page, code, training data, and checkpoint, makes the release useful for inspection, reuse, and further agent-oriented scientific reasoning work.
Meta description: Axion 1.0-35B and 122B use about 18K supervised fine-tuning examples, two lightweight stages, then reinforcement learning for scientific reasoning and tool use.

Tags: Axion 1.0, agent foundation models, scientific reasoning, tool use, open checkpoint
Featured image: Abstract friendly illustration of a small robot reading a molecular model, no real people, no logos, no text.
This is for developers, researchers, and AI teams looking for a compact way to build agent models that can reason about science and use tools. Axion 1.0 introduces Open Agent Foundation Models, Simplified: a post-training approach that keeps the recipe short while aiming for strong scientific reasoning and tool use.
A short recipe with about 18K supervised fine-tuning examples
The Axion 1.0 announcement says the model learns from approximately 18K SFT examples. SFT means supervised fine-tuning, where the model is trained on examples with desired outputs. The recipe uses two lightweight supervised fine-tuning stages, then reinforcement learning. Reinforcement learning here means the model improves through feedback on its own actions rather than only from fixed examples.
The training path is divided into three goals. First, learn to reason by using model-aware construction to create challenging, self-contained scientific problems. Second, learn to investigate through verifiable tasks that require retrieving information and using tools. Third, learn through interaction by applying reinforcement learning over agent trajectories, meaning the full sequence of actions the model takes while working on a task.
35B and 122B aim for competitive scores on six settings
Axion 1.0-35B and 122B are described as achieving competitive results across six evaluation settings: HLE, BioMysteryBench, SEAL-0, HiPhO, and the Olympiad and Research splits of FrontierScience.
For the Axion 1.0-35B reproduction checkpoint, the reported scores are:
- HLE: 52.8
- BioMysteryBench: 35.3
- SEAL-0: 60.4
- HiPhO: 45.9
- FrontierScience Olympiad: 85.0
- FrontierScience Research: 55.0
The strongest listed score is 85.0 on FrontierScience Olympiad, while the 52.8 HLE result is one of the key numbers in the release. The announcement does not provide separate benchmark numbers for the 122B model in the same score list.
Why the Apache 2.0 checkpoint matters
The Axion 1.0-35B reproduction checkpoint is now available under Apache 2.0, a widely used open-source software license. For readers, that makes the checkpoint available for inspection and reuse as a starting point for agent-oriented scientific reasoning and tool-use work.
The release also points to the paper, project page, code, and training data, so readers can trace how the approximately 18K supervised fine-tuning examples and the three learning stages fit together.
Resources to check
- Paper: https://arxiv.org/abs/2611.03233
- Project: https://axion1-0.github.io/
- Code: https://github.com/bytedance/axion1_0
- Training data: https://huggingface.co/datasets/Axion1-0/axion-training-data
- Model (35B reproduction): https://huggingface.co/Axion1-0/Axion-1.0-35B
Comments (0)