The community guide entry for LimonLM Lite Old. The same model as Disintegration Loop Lite, but quantized to fp4 instead of fp8 — making it the only fp4 custom model on the entire Lightshade roster.
Lite Old is the deprecated Lite — the very same model that killed everyone during its LumenBench run. This version runs at fp4 precision.
What is fp4 vs fp8?
fp4 and fp8 refer to the numerical precision used to store the model's weights — fp8 uses 8-bit floating point, fp4 uses 4-bit. Lower precision means the model is smaller and faster, but loses fidelity. Think of it like image compression: fp8 is a high-quality JPEG, fp4 is a heavily compressed one. You can still tell what's in the picture, but fine details — subtle word choices, character voice consistency, nuanced emotional shifts — get blurred. fp4 models tend to repeat themselves more, lose character voices faster, and make blunter creative choices. Lite Old is the only custom model on Lightshade running at fp4. Every other custom model runs at fp8.
Family: LimonLM (Custom)
Base Model: —
Context Window: —
Generation: —
Parameters: —
Architecture: —
Training Finished: —
Release Date: —
Max Token Context: —
Effective Memory: —
Notes on memory behavior: —