Axion 1.0 releases an open agent model training recipe
Axion 1.0 is an open agent foundation model trained in three stages: two lightweight supervised fine-tuning stages followed by reinforcement learning, without dedicated mid-training, direct preference optimization or on-policy distillation. The release includes data construction, generation, filtering, verification, prompts and training code. The paper reports that the 122B model scored 59.0% on the full Humanity’s Last Exam with restricted website access and 66.7% on FrontierScience Research.
The release goes beyond model weights by providing the complete recipe for data construction, verification and training. The paper says these materials enable the development process to be reproduced and extended, offering a practical and reproducible approach to open agentic AI research.
Computer Science > Artificial Intelligence
arXiv:2611.03233 (cs)
[Submitted on 2 Nov 2026 (v1), last revised 3 Nov 2026 (this version, v2)]
Title:Axion 1.0: Open Agent Foundation Models, Simplified
Authors:Yuqi Jia, Yunjie Tian, Jiayu Zheng, Yijun Liang, Jingchen Sun, Mengyu Zhao, Di Fu
View a PDF of the paper titled Axion 1.0: Open Agent Foundation Models, Simplified, by Yuqi Jia and 6 other authors
#Axion 1.0#ByteDance#Qwen3.6-35B-A3B#Qwen3.5-122B-A10B#Duke University#University of Maryland, College Park#University at Buffalo