Ternary Model Playground

Interactive demo for ternary-quantized multimodal models using a HuggingFace-native PTQ workflow.

These models are quantized with ternary-quant using component-aware tritplane3 quantization. Quality retention is benchmark- and model-dependent; see the linked model cards for validation details.

Collection: github.com/Asad-Ismail/ternary-models

How it works: Each model's weights are quantized to ternary values {-1, 0, +1} using progressive tritplane quantization. Vision encoders and text decoders are quantized independently with component-aware calibration. See the ternary-quant paper for details.