Open-weight Kimi K3.1 packs 3 trillion parameters with MoE

Kimi K3.1 is an open-weight multimodal model with 3 trillion parameters and a mixture-of-experts design activating 104 billion per task. It handles text, images and possibly audio, excelling in coding, reasoning and 3D modeling but lagging on science benchmarks and complex 3D rendering.

• 3 trillion parameters, MoE activates 104 billion per task
• Multimodal: text, images and potentially audio
• Strong in coding, reasoning and parametric 3D modeling
• Weaker on science benchmarks and complex 3D rendering