Kimi K2 (Moonshot)

Sobre Moonshot AI

Fundado em 2023 · Beijing, China

Moonshot AI (月之暗面, literally 'dark side of the moon') was founded in March 2023 in Beijing by Yang Zhilin (CEO, Tsinghua and CMU alumnus, co-author of XLNet and Transformer-XL), Zhou Xinyu and Wu Yuxin. The company quickly emerged as one of the four 'AI tigers' of China alongside Zhipu AI, MiniMax and 01.AI. Moonshot's flagship product is Kimi (named after the explorer Kimi Raikkonen), a consumer-facing chatbot launched in October 2023 that became famous in China for very long context handling. Kimi initially launched with a 200K Chinese-character context, expanded to 1M and then 2M tokens via long-context architectural research. Moonshot has raised over $1.3B across multiple rounds, including a $1B round in early 2024 led by Alibaba at a $2.5B valuation, and a subsequent round in late 2024 reportedly valuing the company near $3.3B. Investors include Alibaba, Tencent, Sequoia China, Hongshan and HSG. Kimi K2 was released in July 2025 as the lab's open-weight Mixture-of-Experts flagship, marking Moonshot's move to a more open model strategy alongside its consumer app.

Arquitetura

Sparse Mixture-of-Experts Transformer (Muon optimizer, agentic post-training)

Kimi K2 was released by Moonshot AI in July 2025 as an open-weight Mixture-of-Experts model with 1 trillion total parameters and 32 billion active per token across 384 experts (8 selected per token). The architecture follows DeepSeek-style fine-grained MoE with Multi-head Latent Attention for efficient inference. Kimi K2 was pretrained on 15.5 trillion tokens of multilingual web text, code, books and scientific papers, with a heavy emphasis on Chinese and English. The training run used the Muon optimizer (Momentum Orthogonalized by Newton-Schulz) at unprecedented scale and reportedly improved sample efficiency over AdamW. Moonshot published the MuonClip variant that adds gradient clipping to stabilise Muon at trillion-parameter scale. Post-training emphasised agentic capabilities: Kimi K2 was trained with synthetic and real tool-use trajectories covering multi-step web search, coding, file manipulation and structured tool calling, positioning it as one of the strongest open-weight agentic models. The model supports a 128K context window. Two variants were released: Kimi K2-Base for fine-tuning and Kimi K2-Instruct for general chat and agentic use. Weights ship under a Modified MIT License that requires attribution for very-large-scale commercial deployments.

Parâmetros

1T total, 32B active per token

Contexto

128.000 tokens

Capacidades

• 1T-parameter Mixture-of-Experts with 32B active per token
• Pretrained on 15.5T tokens using the Muon optimizer (MuonClip variant)
• Strong agentic capability: SWE-bench Verified, Terminal-Bench, ToolBench leadership
• 128K context window
• Function calling and parallel tool calls
• Excellent Chinese-English bilingual performance
• Code generation and editing across major languages
• Open weights under Modified MIT License
• Compatible with vLLM, SGLang, llama.cpp, HuggingFace
• Available via Kimi consumer app and Moonshot API
• Best for: agentic workloads, bilingual chat, coding agents, on-prem deployment.

Treinamento & licença

Pretrained on 15.5 trillion tokens of multilingual web text (Chinese and English dominant), code, books and scientific papers using the Muon optimizer with MuonClip stability. Knowledge cutoff is approximately early 2025. Post-training emphasises agentic tool-use trajectories and supervised fine-tuning on curated coding and reasoning data.

Licença: Modified MIT License: open weights, commercial use permitted; very-large-scale deployments require attribution in product UI.

Testes de segurança: Limited published safety evaluations. Filters Chinese political topics in line with regulations; otherwise relatively low refusal rate on Western dual-use queries.