Bytedance argues most AI training phases are unnecessary with new open model
Researchers from TikTok parent Bytedance and several US universities have introduced Axion 1.0, an open AI agent designed to challenge leading commercial systems from OpenAI and Anthropic while drastically simplifying how such models are trained. According to the team's newly published paper, Axion 1.0 rivals top proprietary models on benchmarks after undergoing only three training phases, with the researchers making the model, its training data, and its underlying code freely available.
The project targets a growing complexity problem in the development of autonomous AI agents. Modern AI systems are increasingly expected to research the web independently, run code, and complete multi-step tasks, and developers have typically pursued those capabilities by stacking numerous training phases one after another into deeply nested pipelines.
The researchers argue that this layered approach carries major drawbacks: once multiple stages are combined, it becomes almost impossible to tell which step actually helps, and reproducing the work grows increasingly expensive. By prioritizing carefully curated training data over extra training phases, the Axion 1.0 team aims to show that a streamlined three-step process can match leading commercial systems without that added complexity.
Continue in the app — vote & join in ➔