长鑫科技批量造富:12名核心成员身家过亿,碧桂园却卖飞痛失490亿;哈啰电车骑行中断电落锁致博士十级伤残;月之暗面开源Kimi K3模型
Merged summary
TL;DR - Moonshot AI open-sourced Kimi K3, a 2.8-trillion-parameter MoE model, alongside its weights, technical report, and training infrastructure. The release advances open-model scaling while improving compute efficiency.
- Kimi K3 activates 16 of 896 experts using the Stable LatentMoE framework.
- Architectural, training, and data improvements reportedly deliver about 2.5× better scaling efficiency than Kimi K2.
- MoonEP, FlashKDA, and AgentEnv cover high-performance communication, optimized kernels, and distributed reinforcement-learning environments.
- MoonEP and AgentEnv were newly open-sourced; FlashKDA was already available.
Sources (1)
长鑫科技批量造富:12名核心成员身家过亿,碧桂园却卖飞痛失490亿;哈啰电车骑行中断电落锁致博士十级伤残;月之暗面开源Kimi K3模型
TL;DR - Moonshot AI open-sourced Kimi K3, a 2.8-trillion-parameter MoE model, alongside its weights, technical report, and training infrastructure. The release advances open-model scaling while improving compute efficiency.
- Kimi K3 activates 16 of 896 experts using the Stable LatentMoE framework.
- Architectural, training, and data improvements reportedly deliver about 2.5× better scaling efficiency than Kimi K2.
- MoonEP, FlashKDA, and AgentEnv cover high-performance communication, optimized kernels, and distributed reinforcement-learning environments.
- MoonEP and AgentEnv were newly open-sourced; FlashKDA was already available.