中国AI公司月之暗面于7月27日正式开放Kimi K3完整模型权重，该模型参数规模达2.8万亿，成为全球首个达到此量级的开源模型[1][5][6]。K3在Hugging Face平台发布仅30分钟内获得超4000个点赞，创下该平台增长速度最快纪录[6][9]。模型采用Modified MIT许可证，允许商用、修改和再分发[5]。
月之暗面对K3的商业化采取了分层策略。根据自定义许可证规定，企业及其关联方在任意连续12个月内合计收入超过2000万美元，使用K3进行商业服务需另行签署协议[1]。官方API定价为输入3美元、输出15美元，缓存命中0.30美元[5]。月之暗面表示编程场景缓存命中率可超90%，实际输入成本约为标价的1/10[5]。腾讯云ADP、阿里云百炼等云厂商已在第一时间上线K3调用通道[5]。
然而，K3的开源并未降低技术部署门槛。最低部署配置需要8张高端加速卡，硬件成本约170-240万元，商业级建议配置则需64张以上加速卡，硬件投入约1360-1900万元，全年云服务成本约409-986万元[7]。完整权重文件达1.56TB[7]。
K3的发布在硅谷引发了关于开源与闭源模型的激烈争论。英伟达CEO黄仁勋在7月24日发起公开信支持开源模型，称中国的开源模型"非常优秀"，表现优秀的开源模型就应该被使用[8]。该公开信初期获得25家公司联署，一天内扩至约77家，包括OpenAI、Google、AMD等企业[7]。微软据报正评估将Kimi K3用于Azure及部分Copilot场景[8]。
与此同时，美国前沿模型公司对开源模型表达了保留态度。OpenAI和Anthropic据报已向美国监管机构游说反映对中国开源模型的担忧[3]。Anthropic CEO Dario Amodei发表声明澄清该公司从未主张禁止开放权重模型，但强调了开源模型带来的风险[7]。作为回应，近200家硅谷初创公司联署致信白宫，警告禁止使用中国开源模型"将有数百家美国初创公司瞬间死亡"[9]。
在技术表现上，K3拥有896个专家模块，每个Token仅激活16个，支持图像、视频和100万Token上下文[6][8]。K3在第三方评测中位列全球前茅，在Terminal-Bench 2.1拿到88.3%，超过Claude Fable 5的88.0%；在编程评测SWE-Marathon得到42分，高于GPT-5.6 Sol的39分和Claude Opus 4.8的40分[6]。
此次发布反映了开源模型行业从比拼参数规模转向争夺商业定价权的新阶段[1]。过去三个月中，中国开放模型连续推出三项重大更新，分别针对价格、长任务能力和模型规模对标美国闭源模型[6]。
Moonshot AI released the complete weights of its Kimi K3 large language model on July 27, marking a significant shift in the competitive landscape between open-source and proprietary AI systems [1][5][6]. The model contains 28 trillion parameters, making it the largest open-weight model globally to date [1][5][6]. The company made the weights freely available for download and local deployment on Hugging Face under a modified MIT license, though the release came with a custom licensing agreement specifying that enterprises and their affiliates generating combined revenue exceeding $20 million USD within any consecutive 12-month period must negotiate separate commercial terms to provide services using K3 [1][7].
The decision to open-source such a large model reflects a fundamental business calculation. Moonshot's annualized recurring revenue (ARR) reached $300 million in mid-June, with API income accounting for over 70 percent of total revenue [1]. The average cost per task for K3 stands at approximately $0.94 USD, roughly half that of Claude Opus [1]. On third-party evaluation platforms, K3 achieved approximately 57 points on the Artificial Analysis intelligence index [1]. In specialized benchmarks, K3 scored 42 points on SWE-Marathon, surpassing both GPT-5.6 Sol at 39 points and Claude Opus 4.8 at 40 points [6]. Hugging Face co-founder and CEO Clem Delangue described this as the platform's fastest-growing release to date [6]. Within 30 minutes of release on Hugging Face, K3 received over 4,000 likes and reached the trending charts [6].
However, deploying K3 in production environments remains prohibitively expensive for most users. The model weights total approximately 1.56 TB in size [7][9]. Minimum deployment configurations require either eight AMD MI355X accelerator cards—costing between $1.7 million and $2.4 million—or 64 or more accelerators for commercial-grade service, representing hardware investments of $13.6 million to $19 million, plus cloud service costs ranging from $4.09 million to $9.86 million per year [7]. This hardware barrier means practical access remains limited despite the open release [5][7].
The release triggered intense industry divisions over open-source model policy. On July 24, Nvidia CEO Jensen Huang posted support for open-source models, describing Chinese open-source models as "very good" [8]. A public letter backing open-weight models and American AI leadership subsequently garnered signatures from approximately 50 technology companies, including OpenAI, Google, Microsoft, Meta, IBM, and AMD, though Anthropic notably abstained [6][8][9]. Anthropic CEO Dario Amodei issued a statement clarifying that "Anthropic has never advocated for banning open-weight models," though the company emphasized concerns about certain AI application risks [7][9].
Meanwhile, OpenAI and Anthropic reportedly lobbied regulatory agencies regarding concerns about Chinese open-source models [3]. Nearly 200 Silicon Valley startups sent letters to the White House warning that prohibiting the use of Chinese open-source models "would cause hundreds of American startups to instantly fail" [9]. Microsoft was reported to be evaluating whether to use Kimi K3 for Azure and certain Copilot scenarios, with unconfirmed reports suggesting potential annual savings of up to $600 million [8].
The release represents a strategic pivot from competing on parameter scale alone toward controlling commercial pricing and licensing terms [1]. Major cloud providers including Tencent Cloud ADP and Alibaba Cloud Bailian have already launched K3 calling channels [5]. Prior to the full model release, Moonshot paused new consumer subscriptions on July 20 due to computational resource constraints [1][5].
