WASTE inference engine streams Kimi K3's 2.78T parameters from NVMe
The dependency-free C engine keeps the model trunk in memory, streams activated experts from disk, and uses remaining RAM as a bounded expert cache.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- SCAIL 2 fan video replaces GTA 6 characters with fat versions
- State machine guardrails govern agent tool use in Claude Code, Cursor
- Hollywood-style racing trailer built with ComfyUI, LTX 2.3 & Krea 2
- Ethan Mollick: Microsoft and Google were daring with early AI
- Scoble: AI reads feet via video camera, no LiDAR