Tiny Kimi K3 fits in 700MB, runs locally on potato hardware

Kimi K3 is compressed to 0.18B parameters (0.10B activated), with a size of only 700MB, retaining the full architecture and new attention design, now loadable on normal hardware.

• edge AI
• local inference
• compression
• Kimi K3
• tiny model
• 0.18B
• 700MB