ByteDance 推出 Axion 1.0,仅三个训练阶段即可匹敌 OpenAI、Anthropic 旗舰模型
Bytedance argues most AI training phases are unnecessary and backs it up with a competitive open model
字节跳动带着 Axion 1.0 来了,三个训练阶段就敢对标 OpenAI 和 Anthropic 的旗舰,靠的是精筛数据而不是堆流程。权重、数据、代码全开源,做智能体和研究训练管线的同学可以直接下下来复现,看看他们砍掉了哪些阶段还能保持成绩。
ByteDance 联合多所美国大学发布开源智能体模型 Axion 1.0,在基准测试中与 OpenAI 和 Anthropic 的顶级商业模型表现相当,而其训练流程仅使用三个阶段。研究团队认为,传统训练管线中的多数阶段并不必要,性能提升主要来自精心筛选的训练数据,而非不断增加的训练步骤。模型权重、训练数据和代码均已开放下载。
技术上,Axion 1.0 的核心思路是用数据质量替代流程复杂度:研究者在三个训练阶段内完成全部训练,省略了常规管线中的额外环节,仍能在智能体任务基准上达到商业闭源模型的水平。这一对比直接挑战了主流实验室通过堆叠 SFT、RLHF 等多阶段流程提升能力的做法。
由于模型、数据和代码全部开源,开发者可以复现其训练管线并验证数据筛选策略的贡献,对预算有限的团队和研究机构尤具参考价值。在开源模型普遍落后闭源旗舰几个月的背景下,Axion 1.0 提供了一条更精简的追赶路径。
Bytedance argues most AI training phases are unnecessary and backs it up with a competitive open model
Bytedance and several US universities introduce Axion 1.0, an AI agent that rivals top commercial models from OpenAI and Anthropic on benchmarks after just three training phases. The researchers attribute the gains to carefully curated training data rather than ever more training steps. The model, data, and code are freely available. The article Bytedance argues most AI training phases are unnecessary and backs it up with a competitive open model appeared first on The Decoder .
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Axion 1.0, ByteDance, OpenAI, Anthropic.
Source attribution: curated from Decoder (@Jonathan Kemper).