Moonshot AIがKimi K3のモデル重みを公開、最大規模のオープンウェイトAIを提示
重要ポイント
- •Kimi K3はMoonshot AIの最新大規模言語モデルで、モデル重みと技術レポートがオープンウェイトとして公開された。
- •同モデルは2.8兆パラメータを持ち、圧縮後の重みファイルは約1.4テラバイトで、公開済みオープンウェイトモデルとして最大規模。
- •十分な計算基盤を持つ組織は、Moonshotの許可や専用ホスティングなしに、モデルを導入、改変、商用化できる。
- •MoonshotのKimi K3 API価格は、キャッシュ済み入力100万トークンあたり0.30ドルからで、OpenAIのGPT-5.6やMetaのMuse Spark 1.1より低い。
- •Arena AIはKimi K3 (Max) をAgent Arenaでオープンウェイトモデル首位とし、Frontend CodeおよびText Arenaでも首位に位置づけた。
Moonshot AI, the Chinese AI startup, has released the model weights and technical report for Kimi K3, its latest large language model. The release makes the system available as an open-weight download, allowing developers, enterprises, and government entities with sufficient computational infrastructure to deploy, modify, and commercialize the model without seeking permission from Moonshot or relying on its proprietary hosting infrastructure.
Kimi K3 comprises 2.8 trillion parameters, and the complete weight file is approximately 1.4 terabytes when compressed using MXFP4 quantization, a format designed to reduce storage requirements while preserving most of the model’s inference performance. By both total parameter count and distribution size, it is the largest open-weight model ever released to the public, a milestone that matters because access at this scale can shift more of the competitive debate from who can train the biggest model to who can operate and integrate it most effectively.
The release comes amid rapid price compression across the AI industry. Moonshot’s API pricing for Kimi K3 starts at $0.30 per million tokens for cached input, undercutting offerings such as OpenAI’s GPT-5.6 at $1 per million tokens and Meta’s Muse Spark 1.1 at $1.25 and $4.25 per million tokens for input and output tiers, respectively. By removing the API paywall for organizations capable of self-hosting, the open-weight distribution changes the competitive calculus for commercial laboratories that charge for access to models of comparable scale, including those Moonshot has been accused of using as training references.
ベンチマーク性能と技術評価
According to the latest findings published by the independent evaluation platform Arena AI, Kimi K3 (Max) is positioned at the forefront of open-weight model capability. In the Agent Arena, which evaluates models on millions of real-world, long-horizon agentic tasks, Kimi K3 (Max) achieved a net improvement of 9.75%, surpassing the previous leader GLM-5.2 (Max) at 7.12%. The model secured top rankings across five distinct evaluation signals.
Big update: Among open-weight models, Kimi K3 (Max) is #1 in the Agent Arena with +9.75% net-improvement, surpassing GLM-5.2 (Max) at +7.12%, and landed the #1 spot across 5 signals (see below). Kimi K3 (Max) is also now #1 in open-weight in the Frontend Code (1682 pts) and Text… pic.twitter.com/YATt1LiNYh — Arena.ai (@arena) July 27, 2026
The Agent Arena methodology gives models access to web search, filesystem, and terminal tools to complete complex workflows, including code generation, presentation creation, web research, application development, and document analysis. Arena AI uses causal tracing methodology to measure net improvement, quantifying a model’s performance advantage relative to the average benchmark participant.
Beyond agentic task execution, the model also claimed the leading position among open-weight models in specialized technical domains, scoring 1,682 points in the Frontend Code Arena and 1,485 points in the Text Arena. The results suggest that the model’s scale translates into measurable capability across both autonomous multi-step workflow execution and domain-specific content generation tasks, reinforcing its position within the open-weight ecosystem.