跳到主内容
精选90DAIR.AI(RSS)模型发布/更新多源精选 ×11

月之暗面开源Kimi K3:2.8T参数MoE模型

🤖 AI Agents Weekly: Kimi K3, Thinking Machines Inkling, Agentic Misalignment, Perplexity SPACE, Sunday ACT-2, GPT-Red, and More

原文
推荐理由

国内团队首个3万亿级开源MoE模型,架构创新和Agent能力突出,做模型训练或Agent开发的同学值得关注权重开放后的实测。

In today's issue:

  • Moonshot open-sources Kimi K3
  • Thinking Machines ships Inkling weights
  • Anthropic finds new agent misalignment
  • Perplexity opens SPACE agent runtime
  • Sunday Robotics unveils ACT-2
  • OpenAI launches GPT-Red red teamer
  • Meituan open-sources LongCat-2.0
  • schema harness saturates ARC-AGI-3

And all the top AI dev news, papers, and tools.

Top Stories

Moonshot Open-Sources Kimi K3

Moonshot AI introduced Kimi K3, billed as the first open 3-trillion-class model, with weights due July 27.

  • Scale and architecture: A 2.8T-parameter Stable LatentMoE activating 16 of 896 experts, with a 1M-token context window and native vision in a single model.
  • New attention design: Kimi Delta Attention delivers up to 6.3x faster decoding in million-token contexts, and Attention Residuals add roughly 25% higher training efficiency at under 2% extra compute.
  • Agentic strength: Leads Terminal-Bench 2.1 and scores 67.3 on DeepSWE, with demos of autonomous kernel optimization, compiler development, and vision-in-the-loop coding.
  • Availability: Live via Kimi.com, Kimi Code, and API from about 0.30 dollars to 15 dollars per million tokens, with open weights following on July 27.

Blog

Read more

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

关联讨论

同一事件的更多信源
月之暗面发布新版Kimi模型引发担忧
TechCrunch AI(RSS)原文
Kimi团队再次发布最大中文模型,接近3T参数
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)原文

相似阅读

另一事件,读法相近