colibri

Run frontier open models on hardware you already own.

PlayableOpen source

Description

A dependency-free C engine that streams mixture-of-experts (MoE) weights from disk, allowing massive models like Kimi K3 (2.8T) to run on consumer hardware with limited RAM. It features a unique 'System One' mode for rapid, calibrated decision probabilities and supports a broad suite of open models including DeepSeek, Qwen, and GLM.

Descriptions, tags, and model credits may be AI-generated or inferred from public sources and can be incomplete or wrong. Learn more.

Details

Why this is interesting: Impressive technical feat: enables running multi-trillion parameter MoE models on consumer hardware by disk-streaming…
Attributed to: Vincenzo Fornaro@JustVugg
Built with: GLiNER2.5-Decide Clef OLMoE Kimi K3 GLM-5.2 V4
Links: justvugg.github.io
View GitHub repo
Tags: #inference #moe #c-engine #open-source

Discovered: X
Listed: Added 2h ago
Build: Strong
Models: Strong

colibri

Description

How it grew

colibri

Description

How it grew