88Z AI, 28 July — Moonshot AI has released the Kimi K3 model weights, technical report, and three key infrastructure technologies. The model features a 2.8 trillion parameter MoE architecture with native visual understanding and a one-million token context window, offering significantly improved efficiency. Additionally, MoonEP, FlashKDA, and AgentEnv have been open-sourced, covering critical links from training communication to distributed reinforcement learning environments.
At a glance
- Kimi K3 is a 2.8 trillion parameter MoE model with native visual support and a one-million token context, offering 2.5x efficiency gains.
- MoonEP, FlashKDA, and AgentEnv are open-sourced, covering key technical links from training communication to distributed RL environments.
The update
Moonshot AI has officially released the model weights and technical report for Kimi K3, along with three key infrastructure technologies. Kimi K3 is a mixture-of-experts model with 2.8 trillion parameters, featuring native visual understanding and support for a context window of up to one million tokens. Its scaling efficiency has improved by 2.5 times compared to the previous generation. The open-sourced infrastructure includes the high-performance communication library MoonEP, the FlashKDA operator, and the AgentEnv sandbox system. These technical components cover critical links from model training communication to distributed reinforcement learning environments, aiming to provide developers with a more comprehensive technical support system.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 月之暗面KimiSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.