Kimi K3: Open-Weight Trillion-Scale MoE for Agentic Intelligence

Authors: Moonshot AI (2026)

Domains

Mixture of ExpertsMultimodalApplications

TLDR (English)

A MoE model with ~2.8T total parameters and ~100B activated, at release the largest open-weight model ever. Natively multimodal with ~1M-token context, focused on agentic capability—one of the landmark 2026 engineering efforts in which open weights caught up with the closed frontier.

TLDR（中文）

约 2.8T 总参数、百亿级激活参数的 MoE 模型，开放权重后成为当时规模最大的开源权重模型，原生支持视觉与约 1M token 上下文，主打 Agentic 与多模态能力，是 2026 年开源权重追平闭源前沿的标志性工程之一。

Appears in These Articles

• 预训练与 Scaling Law：模型怎么学
• Pretraining and Scaling Law: How Models Learn

Co-cited Papers

These papers appear in the same articles as this one

• Language Models are Few-Shot Learners Tom Brown et al. (2020)
• PaLM: Scaling Language Modeling with Pathways Aakanksha Chowdhery et al. (2022)
• DeepSeek-V4 Technical Report DeepSeek-AI (2026)
• Training Compute-Optimal Large Language Models Jordan Hoffmann et al. (2022)

Related Papers

Other papers in the same domain

• Agent World Model: Scaling Agentic RL with Synthetic Verifiable Environments Agent World Model Team (2026)
  ApplicationsReasoning
• DeepSeek-V4 Technical Report DeepSeek-AI (2026)
  ArchitectureLong ContextMixture of Experts
• Model Context Protocol (MCP) Anthropic (2024)
  Applications
• DeepSeek-V3 Technical Report DeepSeek-AI (2024)
  ArchitectureMixture of Experts
• SWE-bench: Can Language Models Resolve Real-World GitHub Issues? Carlos E. Jimenez et al. (2024)
  EvaluationApplications
• Mixtral of Experts Albert Q. Jiang et al. (2024)
  Mixture of ExpertsArchitecture