北京初创公司月之暗面（Moonshot）发布了新一代AI模型Kimi K3，其性能在多项基准测试中逼近甚至超越美国顶尖模型Claude和ChatGPT，引发美国科技行业的广泛关注[1]。
性能表现与评测结果
Kimi K3采用2.8万亿参数架构[2]，在Arena的"前端编码能力"排名中位居榜首[1]。根据Artificial Analysis Intelligence Index的AA-Briefcase知识工作基准测试，该模型获得1543分的Elo评分，仅次于Claude Fable 5的1574分，相比前代Kimi K2.6提升了727分[2]。分析师Anastasios Angelopoulos称这可能是"今年最大的发布"，并指出开源中国模型正在超越封闭的美国模型[1]。
在其他指标方面，Kimi K3的Rubric通过率为51%，分析质量Elo为1754，均处于业界领先水平[2]。不过，该模型的运行成本相对较高，平均每任务成本为10.57美元，平均耗时56.4分钟，约为Claude Fable 5的2.5倍[2]。
定价与商业竞争
Kimi K3的定价为每100万个input/output tokens各3美元和15美元[2]，仅为OpenAI GPT-5.6 Sol模型一半[1]。这种定价策略使其在成本敏感的市场中具有竞争力。
蒸馏技术与知识产权争议
美国指控月之暗面通过蒸馏技术获取Anthropic的Fable模型用于K3开发[3]。根据指控，月之暗面建立了内部平台进行大规模蒸馏，并采用多种方法规避检测[3]。此外，该公司购置了配备GB300芯片的服务器，并在泰国访问GB300资源用于AI模型训练[3]。美国官方表示，大规模、隐蔽的工业蒸馏盗取美国专有技术是不可接受的[3]。Anthropic曾指控包括DeepSeek和MiniMax在内的多家中国AI公司同样利用蒸馏技术非法提取Claude的能力[1]。
与此同时，旧金山初创公司Anysphere承认其热门编码工具Cursor基于月之暗面的K2.5模型构建[1]。
创始人背景
月之暗面联合创始人杨志林于2019年在卡内基梅隆大学获得博士学位[1]，其教育背景引起了前同事的关注。
Beijing-based Moonshot AI has unveiled Kimi K3, a large language model with 2.8 trillion parameters that demonstrates performance comparable to leading American systems, marking a significant milestone in China's artificial intelligence development [2]. The release has caught the attention of the US technology industry and sparked concerns about the narrowing gap between Chinese open-source AI and American closed-source models [1].
On industry benchmarks, Kimi K3 ranks second on the AA-Briefcase knowledge work evaluation with an Elo score of 1543, trailing only Claude Fable 5's 1574 [2]. The model represents a substantial improvement from its predecessor Kimi K2.6, gaining 727 points on the same benchmark [2]. It also achieved top rankings in front-end coding ability on the Arena leaderboard [1]. Analysis quality on the new model reaches an Elo score of 1754, matching Fable 5's 1744, though presentation quality trails GPT-5.6 Sol at 1471 compared to that model's 1660 [2].
Kimi K3 carries a significant operational cost of $10.57 per task with an average processing time of 56.4 minutes, making it among the most expensive and slowest models currently available [2]. The model is priced at $3 and $15 per million input and output tokens respectively [2]. Analyst Anastasios Angelopoulos described the launch as potentially "the biggest release of the year," noting that open-source Chinese models are now surpassing closed American counterparts [1].
However, the release has drawn allegations from US companies. Anthropic has accused Moonshot, along with DeepSeek and MiniMax, of illegally extracting its model capabilities through distillation techniques [1]. According to US officials, Moonshot built internal platforms for large-scale distillation and employed multiple access methods to evade detection while developing Kimi K3 [3]. The company also acquired servers equipped with GB300 chips and accessed GB300 resources in Thailand for AI model training [3]. The US government has stated that large-scale, covert industrial distillation of American proprietary technology is unacceptable [3].
The controversy extends to commercial partnerships. San Francisco-based startup Anysphere acknowledged that Cursor, its popular coding tool, was built on Moonshot's K2.5 model [1]. Elon Musk's SpaceX has announced plans to acquire Cursor for $6 billion later this year [1].
Moonshot's co-founder Yang Zhilin earned a doctoral degree from Carnegie Mellon University in 2019 [1].