A new open agent foundation model, Axion 1.0, has been introduced, featuring strong agentic capabilities developed through a simplified three-stage training process. This method involves lightweight supervised fine-tuning and reinforcement learning, notably omitting more complex techniques like DPO or OPD. The model's training recipe, data construction, and code are fully open-sourced, allowing for reproduction and further development. AI
IMPACT Provides a reproducible and open path for developing advanced agent foundation models, potentially accelerating research and application in agentic AI.
RANK_REASON Frontier-lab model release with system card and open-source recipe. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →