阶跃星辰发布 Step 系列

2023 年 11 月，阶跃星辰发布 Step-1，一款千亿参数级别的语言大模型。彼时国内大模型创业方兴未艾，多数新势力选择在通用对话或垂直场景里找位置，而阶跃星辰创始人姜大昕给公司定下的路线更加激进：做大参数、做多模态、做自研。在「算力昂贵、竞争激烈」的行业共识下，这几乎是一场豪赌。

阶跃星辰的背景并不寻常。姜大昕是前微软亚洲研究院的 NLP 方向资深研究者，团队核心成员多来自微软、字节等头部 AI 团队。这种「学术 + 工程」的底子，让阶跃星辰在一众创业公司里显得更「硬核」。Step-1 之后，2024 年发布的 Step-2 直接把参数规模推到万亿级别，成为当时少数能在参数规模上与海外前沿模型对齐的国产模型之一。

Step 系列的产品化路径也在演化。早期以语言模型为主，随后逐步转向多模态与 Agent 场景——文本、图像、视频、语音能力的整合，加上面向开发者的 API 与部分开源权重。阶跃星辰希望证明的不只是「我们能造大模型」，还有「我们造的大模型能用在真实产品里」。2025-2026 年，Step 系列持续迭代，成为「大模型六小龙」里技术路线最鲜明的一家。

不过，重参数路线也意味着重负担。训练万亿参数模型的算力成本、对商业化的现实压力、以及开源与闭源之间的权衡，都是阶跃星辰必须面对的问题。它选择了一条更难走的路——不靠包装和营销，靠技术实力说话。

回看阶跃星辰的 Step 系列，它是中国大模型创业的一个极端样本：把几乎所有筹码押在「自研能力」上。当行业后来陷入参数内卷又转向效率竞赛时，阶跃星辰的万亿参数路线提供了一次真实的技术试错——它证明中国团队有能力做大规模前沿模型，也把「为什么做、怎么做、值不值得做」的问题留给了整个行业。

In November 2023 StepFun released Step-1, a language model at the hundred-billion-parameter scale. At the time China's model-startup wave was just getting going, and most new entrants looked for a position in general chat or vertical scenarios. StepFun's founder Jiang Daxin set a more aggressive route: big parameters, multimodal, full self-research. Against the industry consensus of "expensive compute, brutal competition," this was nearly a gamble.

StepFun's background was unusual. Jiang was a senior NLP researcher at Microsoft Research Asia, and core team members came from Microsoft, ByteDance, and other top AI groups. This "academic plus engineering" foundation made StepFun seem more hardcore than most startups. After Step-1, Step-2 in 2024 pushed parameters to the trillion scale, making it one of the few domestic models that could align with overseas frontiers on parameter count.

Step's productization also evolved. It started with language models, then shifted toward multimodal and agent scenarios—integrating text, image, video, and speech, plus developer APIs and some open weights. StepFun wanted to prove not just "we can build foundation models" but "the models we build work in real products." Through 2025-2026 the Step line kept iterating, becoming the most technically distinctive of the "six little dragons."

But the heavy-parameter route means heavy burdens. The compute cost of training trillion-parameter models, the pressure to commercialize, and the balance between open and closed all confront StepFun. It chose a harder road—winning on technical substance, not packaging and marketing.

Looking back at StepFun's Step line, it is an extreme sample of Chinese model entrepreneurship: betting nearly everything on "self-research capability." When the industry later tired of parameter inflation and turned to efficiency competition, StepFun's trillion-parameter route provided a real technical experiment—proving Chinese teams can build frontier-scale models, while leaving the whole industry to answer "why do it, how, and is it worth it."