SmolVLM2-2.2B-Instruct — Ternary Quantized (tritplane3)

Ternary-quantized version of HuggingFaceTB/SmolVLM2-2.2B-Instruct using ternary-quant.

Compact VLM designed for edge deployment, now even smaller with ternary quantization.

Model Specifications

Property	Value
Base Model	HuggingFaceTB/SmolVLM2-2.2B-Instruct
Parameters	2.2B
Architecture	VLM (image + text)
Quantization	tritplane3 (169 layers, 10.92 effective bits)
Vision Encoder	FP16 (preserved)
Compression	1.47x
Avg Reconstruction Error	0.1236
License	Apache 2.0

Size Comparison

Method	Size	VLM Support
FP16 (original)	~4.4 GB	Yes
Ternary tritplane3	1.8 GB	Yes

No GGUF alternative exists for SmolVLM2.

Quality Verification

Validated during quantization (collapse score: 0.009 — excellent):

Test	Output
Image description (demo)	"A yellow circle with a diagonal line through it" (correct)
"What is machine learning?"	Correct, detailed explanation of ML, algorithms, training
"Explain gravity"	Accurate one-sentence explanation

Memory Requirements

Runtime	Min Memory	Hardware
`cached` (CPU)	~4 GB RAM	Any
`metal` (Apple Silicon)	~3 GB unified	M1+
`cached` (CUDA)	~3 GB VRAM	Any NVIDIA GPU

Ideal for edge deployment — runs on devices with 4 GB RAM.

Quickstart

pip install ternary-quant

from ternary_quant.inference import load_ternary_model

model, processor = load_ternary_model(
    "AsadIsmail/SmolVLM2-2.2B-Instruct-ternary",
    runtime_mode="cached", device="auto"
)

inputs = processor(text="Describe this image", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128)
print(processor.decode(outputs[0], skip_special_tokens=True))