THE ESSENTIALS IN 30 SECONDS

  • Axion-1.0-35B is a mixture-of-experts model with about 35B total and 3B active parameters, released under Apache 2.0.
  • The authors publish the paper, training code and an Apache-2.0 dataset of 18,363 fine-tuning trajectories and 484 reinforcement-learning tasks.
  • On their own evaluations, the team says the 35B model beats or ties Gemini 3.1 Pro on six agentic benchmarks; these results are not independently verified.

A group of researchers led from ByteDance has published Axion 1.0, an "open agent foundation model" meant to reason over scientific problems and call tools. The first checkpoint on Hugging Face, Axion-1.0-35B, is a mixture-of-experts model with roughly 35 billion parameters in total, of which about 3 billion are active per token. It is licensed under Apache 2.0. The accompanying paper went up on arXiv on November 2 and was revised the next day.

What is released

The team says it is publishing the full recipe needed to rebuild the model: BF16 weights, training code hosted under ByteDance's GitHub organisation, and the training set itself. The dataset, also under Apache 2.0, holds 18,847 rows: 10,819 scientific-reasoning conversations in which the solver may run Python but not search, 7,544 verified tool-use trajectories built from public databases, and 484 prompts with reference answers for reinforcement learning.

The project page lists seven authors. Four are at ByteDance, including project lead Yunjie Tian; the others are affiliated with Duke University, the University of Maryland and the University at Buffalo. The page does not name a specific ByteDance team.

How it was trained

The paper argues for a deliberately short pipeline: two light supervised fine-tuning stages followed by reinforcement learning, with no mid-training, DPO or on-policy distillation. The first stage generates hard problems and reinforces the base model's own successful strategies; the second teaches tool use on tasks whose answers can be checked automatically. Reinforcement learning then targets long-horizon work with tools. Neither the model card nor the project page names the base model, although a Hugging Face tag points to a Qwen 3.5 MoE architecture.

Claimed results and caveats

The authors report results on six settings, including Humanity's Last Exam (52.2), SEAL-0 (62.2) and FrontierScience Research (56.7). In their table, the 35B model scores above Gemini 3.1 Pro, which the team evaluated itself, on five of them and ties on the BioMysteryBench subset. The project page notes that Axion figures are best single runs (best of five for BioMysteryBench) and that baseline settings may differ across sources, so the comparison should be read with care. A larger 122B variant is described on the project page, but only the 35B weights are linked. No independent evaluation was available at the time of writing.