독서리서치레이더투자 프레임워크
로그인 / 회원가입
← 읽기로 돌아가기
뉴스주제 · 月之暗面Kimi模型系列중대2 분 소요DisconfirmAI 편집부

Moonshot AI Releases Open Weights for Kimi K3.1 Pro with $20M Commercial Threshold

Moonshot AI has officially released the open weights for its 2.8-trillion parameter model, Kimi K3.1 Pro, on Hugging Face. The custom license allows free commercial use but requires a separate agreement for Model-as-a-Service providers exceeding $20 million in annual revenue.

2026-11-05 16:34 GMT+8

Chinese AI company Moonshot AI officially released the open weights for its large language model, Kimi K3.1 Pro, on the Hugging Face platform on November 5. Previously, the company had launched the model via API on October 23, promising to release the weights by early November. This release fulfills that promise, allowing developers to locally deploy this Mixture-of-Experts (MoE) model with 2.8 trillion total parameters.

Kimi K3.1 Pro is based on the Kimi K3 architecture released in July, further trained with larger computational resources. The model activates 104 billion parameters and supports a context window of 1 million tokens, handling both text and image inputs. Architecturally, it replaces full attention layers with Gated DeepSeek Sparse Attention (Gated DSA) to optimize long-context processing. Moonshot recommends vLLM, SGLang, and TokenSpeed as inference engines.

According to vendor self-reported data, the model scored 65.7% on the Terminal-Bench 4.0 coding agent test, slightly below Claude Opus 5.5 (66.4%) but above GPT-6 Astra (58.2%). In 3D production tasks, it achieved 92.7% accuracy on the BenchCAD benchmark when using tools, also trailing Opus 5.5 and GPT-6 Astra. Notably, to reduce security risks, Moonshot intentionally removed cyber-attack-related data during training, resulting in significantly lower offensive code generation capabilities compared to unfiltered checkpoints.

The release uses a custom license named "Kimi K3.1-Pro License." This license generally permits free use, reproduction, modification, redistribution, sale, and fine-tuning. However, if enterprises provide the model as a service (Model as a Service) to third parties, and their combined associated companies' total revenue exceeds $20 million in any consecutive 12-month period, they must sign a separate commercial contract with Moonshot AI. Additionally, commercial products with more than 100 million monthly active users or over $20 million in monthly revenue must prominently display "Kimi K3.1-Pro" on their user interface.

로그인 / 회원가입
독서리서치레이더투자 프레임워크

출처

  • pc.watch.impress.co.jpPC Watch↗
주제

月之暗面Kimi模型系列

月之暗面(Moonshot AI,中国)于2026年10月发布Kimi K3.1 Pro模型,并于11月5日在Hugging Face正式开源其2.8万亿参数权重。该模型采用MoE架构,支持百万上下文,在编码基准Terminal-Bench…

4 단계4 건업데이트됨 2026-11-07 04:12

현재까지의 과정

  1. 2026-10-23月之暗面在API平台发布Kimi K3.1 Pro及Highspeed版本,承诺11月4日前发布完整模型权重。Simon Willison报道其定价与GPT-6.1 Sol持平,低于Claude Opus 5.5。simonwillison.net ↗
  2. 2026-10-24官方博客公布详细基准数据:Terminal-Bench 4.0得分65.7%(略低于Claude Opus 5.5的66.4%),新增3D资产生成能力,BenchCAD得分92.7%。kimi.ai ↗
  3. 2026-11-04vLLM团队宣布为Kimi K3.1 Pro提供首日支持,通过稀疏注意力(DSA)和打包KV缓存优化,在NVIDIA、AMD等硬件上实现高效长上下文推理。vllm.ai ↗

아직 해결되지 않은 문제

  • 自定义许可证中关于‘相关公司’营收合并计算的具体界定标准是否清晰?
  • 稀疏注意力机制在实际生产环境中对非vLLM框架的兼容性表现如何?
  • 开源权重后,社区微调版本是否会突破原模型在特定领域(如网络安全)的能力限制?

이 스레드의 다른 내용

  • vLLM Delivers Day-0 Support for Kimi K3.1 Pro: Sparse Attention and Packed KV Cache Boost Long-Context Inference Efficiency2026-11-05
  • Moonshot AI Releases Kimi K3.1 Pro: Coding Benchmark Nears Claude Opus 5.5, Adds 3D Asset Generation2026-10-24
전체 주제 읽기 →
리서치 계속하기 →