A new open agent model built for scientific reasoning and tool use is now on Hugging Face. Called Axion 1.0-35B, the repository publishes a reproduction checkpoint of the version described in the paper Axion 1.0: Open Agent Foundation Models, Simplified, released under the Apache 2.0 license.

Psst!Add AlexTech.ai asPreferred Source on Google

How it is trained

The training recipe combines about 18,000 supervised examples in three stages. The first teaches reasoning on self-contained scientific problems built with model-aware feedback. The second teaches investigation: the model works on verifiable tasks that require external information and tool use. The third is reinforcement learning over agent trajectories, to refine reasoning and tool interaction. The paper stresses that the setup is deliberately simple: two lightweight supervised fine-tuning stages followed by reinforcement learning, without dedicated mid-training or preference optimization.

According to the model card, it is a Mixture of Experts architecture with about 35 billion total parameters, 3 billion active, in BF16 precision. These figures come from the card, which is the single source for them.

The results

The card's table compares paper scores with those of the released checkpoint. On Humanity's Last Exam (HLE) the reproduction scores 52.8%, against 52.2% in the paper. On SEAL-0 it drops to 60.4% (the paper reports 62.2%), while on HiPhO it rises to 45.9% (45.7% in the paper). On FrontierScience Olympiad it reports 85.0%, against 86.0% in the paper, and on FrontierScience Research 55.0% against 56.7%. On BioMysteryBench the figure is identical, 35.3%.

Comparison with other models in the table is more mixed. On HLE, Axion sits below Claude Opus 5 (63.6%), Claude Fable 5 (63.8%), GPT-6 Astra (57.2%) and Muse-Spark-1.0 (58.0%). On SEAL-0 it is above GPT-5.5 (42.3%) and Gemini 3.1 Pro (53.2%) as reported. These are the team's own numbers and have not yet been checked by independent evaluations.

official image of Axion 1.0

Image: arXiv — https://arxiv.org/abs/2611.03233

How to use it

The card describes serving with vLLM on eight GPUs, with parallelism and context length to be adjusted to available memory. Agent evaluations that call tools need a harness that executes those calls, described in the inference guide of the code repository published on GitHub by ByteDance.

The repository also links the training dataset on Hugging Face, which holds about 18.8 thousand examples. The model has a paper page on Hugging Face, alongside the recently published arXiv paper.