|
中国Moonshot Alは7月27日(現地時間)、同社の最新モデル「Kimi K3」のモデルウェイトと技術レポートを公開しました。Kimi K3は、Moonshot AI社が開発した2.8兆パラメーターを誇る世界最大級のオープンウェイトモデルで、KDA(Kimi Delta Attention)という独自の構造を採用することで、100万トークンを超える超長文処理のコストを劇的に削減しています。性能面ではプログラミングや論理推論で商用モデルに匹敵する一方、1.56TBという巨大なサイズゆえに、運用には大規模なGPUクラスタが必要となるインフラ上の課題も指摘されています。ライセンス体系は、一般的な企業利用は無償とする一方で、一定以上の収益やユーザー数を持つ大規模事業者に特定の義務を課す独自形式を採用しています。日本国内でも「NVIDIA B300」を8基使ったシングルノード環境でのデプロイ報告が上がっており、Perplexityでも米国サーバーでのKimi K3の提供が開始されるなど、使用が広がっているようです。 Kimi K3のモデルウェイトと技術レポートの公開内容、Kimi K3使用に関する評価、評判などを生成AIに調べさせましたので、ご参照ください。なお、生成AIによる調査・分析結果は、公開された情報からだけの分析であり、必ずしも実情を示したものではないこと、誤った情報も含まれていることについてはご留意されたうえで、ご参照ください。 Model Weights and Technical Report for "Kimi K3" Released On July 27 (local time), the Chinese AI company Moonshot AI released the model weights and technical report for its latest large language model, Kimi K3. Kimi K3 is one of the world's largest open-weight models, featuring 2.8 trillion parameters. It adopts a proprietary architecture called KDA (Kimi Delta Attention), which dramatically reduces the computational cost of processing ultra-long contexts exceeding one million tokens. In terms of performance, Kimi K3 delivers programming and logical reasoning capabilities comparable to those of leading commercial models. At the same time, its enormous size—approximately 1.56 TB—poses significant infrastructure challenges, requiring large-scale GPU clusters for deployment and operation. The licensing framework is also distinctive. While the model is generally available free of charge for typical enterprise use, it imposes specific obligations on large-scale organizations that exceed certain revenue or user thresholds. In Japan, there have already been reports of successful deployment on a single-node system equipped with eight NVIDIA B300 GPUs. In addition, Perplexity has begun offering Kimi K3 through its U.S.-based servers, suggesting that adoption of the model is steadily expanding. I asked a generative AI system to conduct an in-depth investigation into the contents of the released Kimi K3 model weights and technical report, as well as the model's evaluations, user feedback, and industry reception. I hope you find the analysis informative. Please note, however, that the investigation and analysis were generated by AI based solely on publicly available information. As such, they may not necessarily reflect the actual situation and may contain inaccuracies. I therefore encourage readers to review the findings with appropriate caution. Your browser does not support viewing this document. Click here to download the document. Your browser does not support viewing this document. Click here to download the document. Your browser does not support viewing this document. Click here to download the document. Your browser does not support viewing this document. Click here to download the document. Your browser does not support viewing this document. Click here to download the document. Your browser does not support viewing this document. Click here to download the document. Your browser does not support viewing this document. Click here to download the document. Your browser does not support viewing this document. Click here to download the document. Your browser does not support viewing this document. Click here to download the document.
0 Comments
Leave a Reply. |
著者萬秀憲 アーカイブ
April 2026
カテゴリー |
RSS Feed