商业产业 TechCrunch AI Fri, 11 Sep 2026 正文 ok 228 words

Kimi 开发商月之暗面瞄准 20 亿美元年化营收

英文标题:Kimi-maker Moonshot AI targets $2B in annual revenue

中国 AI 实验室月之暗面(Moonshot AI)计划在今年年底前实现 20 亿美元年化营收,较其 8 月披露的营收运行率翻倍。其开源权重 K3 模型自今年夏季发布以来表现突出,OpenRouter 数据显示,该平台上的 K3 模型日生成 Token 数最高可达 3000 亿。尽管其营收规模仍明显低于 OpenAI 和 Anthropic,且开放权重模式的利润率较低,但这一目标表明开放权重 AI 模型仍具备可观的商业化潜力。与此同时,Anthropic 指控月之暗面长期通过模型蒸馏获取 Claude Opus 输出,使其模型开发实践面临争议与潜在法律风险。
打开原文相关 2 · 机会 2
media3.1 KB media389 B media91.0 KB media208.5 KB media55.0 KB media202.2 KB media34.7 KB media43.8 KB media61.4 KB media46.5 KB media20.8 KB media208.5 KB media248 B media116.4 KB

中文

中国最知名的 AI 实验室之一月之暗面(Moonshot AI)认为,它可以将广受欢迎的开放权重模型转化为显著的销售增长。周五,彭博社报道,该实验室的目标是在今年年底前实现 20 亿美元年化营收,是该公司 8 月披露营收运行率的两倍。这是一个进取的目标,反映出该公司 K3 模型自今年夏季发布以来所取得的成功。

尽管 K3 的使用数据在近几个月略有下降,OpenRouter 数据目前显示,系统中 K3 模型每天生成的 Token 数最高可达 3000 亿。

月之暗面的预计营收规模仍远低于 OpenAI 和 Anthropic;近期报道称,两者的年化营收分别达到400 亿美元和650 亿美元。由于月之暗面的模型权重可自由获取,该公司的利润率远低于其闭源权重竞争对手。不断上调的预测表明,开放权重 AI 模型仍然能够赚钱,尽管其盈利能力不及闭源权重的前沿模型。

不过,月之暗面的模型开发实践仍存在争议,甚至可能涉嫌违法。本周早些时候,Anthropic 指控该公司长期开展模型蒸馏活动:其将近 30 万个来自 Kimi 的请求直接路由至 Claude Opus,实际上以 Opus 替代 Kimi 自身模型提供服务。Anthropic 指称,月之暗面总共收集了超过 2300 万条 Anthropic 模型回复,用于其训练。

English / 原文

One of China’s most prominent AI labs, Moonshot AI, believes it can turn its popular open-weight model into serious sales growth. On Friday, Bloomberg reported that the lab is targeting $2 billion in annualized revenue by the end of the year, double the company’s reported revenue run rate for August. It’s an aggressive goal that reflects the success of the company’s K3 model since its release this summer.

While K3’s usage figures have declined slightly in recent months, OpenRouter data currently shows as many as 300 billion tokens being generated each day by K3 models on the system.

Moonshot’s projected revenue is still dwarfed by that of OpenAI and Anthropic, which recent reports put at $40 billion and $65 billion, respectively. Because Moonshot’s model weights are freely available, the company has far lower margins than its closed-weight competitors. The rising projections show there’s still money to be made from open-weight AI models, even if they’re not as lucrative as closed-weight frontier models.

Still, Moonshot’s model development practices remain controversial — if not downright illegal. Earlier this week, Anthropic accused the company of a long-running model distillation campaign that routed nearly 300,000 requests from Kimi directly to Claude Opus, effectively serving Opus in place of Kimi’s own models. In total, more than 23 million responses were collected from Anthropic models for use in Moonshot’s training, the company alleged.