PulseAugur
EN
LIVE 03:38:07
中文(ZH) 月之暗面Kimi K3正式开源

Moonshot AI open-sources Kimi K3 model with 2.8T parameters and 1M context

Moonshot AI has officially open-sourced its Kimi K3 model, releasing its model weights and technical report. The company also made available the key infrastructure technologies, MoonEP, FlashKDA, and AgentEnv, used for training Kimi K3. This new model is a 2.8 trillion parameter Mixture-of-Experts (MoE) model with native visual understanding capabilities and a 1 million token context window, available for free download and deployment by the public. AI

IMPACT This open-source release of a large-scale MoE model with advanced context window capabilities could accelerate research and development in the AI community.

RANK_REASON Frontier-lab model release with system card and infrastructure components. [lever_c_demoted from frontier_release: ic=1 ai=1.0]

Read on 36氪 (36Kr) →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

Moonshot AI open-sources Kimi K3 model with 2.8T parameters and 1M context

COVERAGE [1]

  1. 36氪 (36Kr) TIER_1 中文(ZH) ·

    Moon's Dark Side Kimi K3 Officially Open-Sourced

    36氪获悉,7月27日晚,月之暗面发布Kimi K3的模型权重、技术报告,并开源支撑Kimi K3模型训练的关键Infra技术:MoonEP、FlashKDA和AgentEnv。据介绍,Kimi K3是一个拥有2.8万亿参数的混合专家(MoE)模型,具备原生视觉理解能力,并支持100万token的上下文窗口。月之暗面表示,现在每个人都可以下载并部署Kimi K3模型,无论是用于内部研发,还是嵌入到面向终端用户的产品中,均可自由使用。