CANVAS 2D UNAVAILABLE
This demo draws a live quantization lab on a canvas.
Idea in one line: quantization rounds many FP16 weight values onto a few
representable INT levels, cutting memory roughly 2x to 4x while adding a small, measurable approximation error.

Sampling 4096 weights

PROTRAILBLAZER

Quantization · Bits vs Fidelity

BITS / WEIGHT16.00

UNIQUE LEVELSFLOAT

MODEL WEIGHTS7B

WEIGHT MEMORY14.00 GB

COMPRESSION1.00x

MEAN ABS ERROR0.0000

COSINE SIM1.00000

Mode

Group size 128

Dist

Model

Matrix

Morph 0%

FP16 MASS QUANT LEVELS |ERROR|

Simplified on purpose. Round-to-nearest symmetric int on a seeded synthetic 64×64 tensor. Real quality depends on architecture, calibration, method (GPTQ, AWQ, NF4), kernels, activation and KV precision, and the task. Bit count alone does not predict usability. Memory counts weights plus group scales only. Downstream numbers are simulated, not a real LLM.