月之暗面推出的大模型K3凭借2.8万亿总参数成为全球参数最大的模型[1]。在多项性能测试中，K3表现突出，在OmniDocBench推理测试中达到91.1分，超过Claude Fable 5的89.8分[1]；在Arena前端开发榜上以1677分登顶[1]。在成本效益方面，K3单任务成本约0.94美元，约为Claude Opus的一半[1]。
技术层面，K3采用MoE稀疏激活架构，稀疏激活率为1.79%（16/896），相比K2整体训练效率提升约2.5倍[1]。月之暗面的差异化策略体现在定价与全球拓展上：国内API定价为每百万token 2.3美元，显著高于国内竞品[1]，而在国际市场上则具有价格竞争优势。Kimi产品已进入200多个国家和地区，海外市场付费用户和API收入增长了400%[1]。
融资方面，月之暗面的估值大幅上升，从2024年12月的43亿美元升至2025年5月的200亿美元，6月新一轮融资投前估值达315亿美元[1]。月之暗面创始人杨植麟曾为Transformer-XL、XLNet的核心作者[1]。
Moonshot AI has unveiled K3, a large language model boasting 28 trillion parameters, making it the world's largest model by parameter count [1]. The model demonstrates strong performance across multiple benchmarks, achieving 91.1 points on the OmniDocBench reasoning test, surpassing Claude Fable 5's 89.8 score [1], and ranking first on the Arena front-end development leaderboard with a score of 1677 [1].
K3's computational efficiency represents a significant advancement in the field. The model employs a mixture-of-experts sparse activation approach with an activation rate of 1.79%, improving overall training efficiency by approximately 2.5 times compared to its predecessor K2 [1]. Per-task inference costs amount to approximately $0.94, roughly half the $1.80 cost of Claude Opus [1]. However, K3's domestic API pricing is set at $2.3 per million tokens, notably higher than competing Chinese models [1].
The launch reflects Moonshot AI's rapid commercial expansion. The company's valuation has surged from $4.3 billion in December 2024 to $20 billion in May 2025, with a pre-money valuation of $31.5 billion for its June 2025 funding round [1]. Kimi, the company's consumer-facing product, has expanded to over 200 countries and regions, with paid users and API revenue in overseas markets growing by 400% [1]. Moonshot AI's founder Yang Zhilin is a core contributor to the foundational Transformer-XL and XLNet architectures [1].
