|
|
--- |
|
|
license: apache-2.0 |
|
|
base_model: |
|
|
- HuggingFaceTB/SmolVLM2-2.2B-Base |
|
|
pipeline_tag: image-text-to-text |
|
|
library_name: transformers |
|
|
--- |
|
|
|
|
|
**SmolVLM2‑2.2B‑Base Quantized** |
|
|
|
|
|
--- |
|
|
|
|
|
### 🚀 Model Description |
|
|
|
|
|
This is a **quantized version** of **SmolVLM2‑2.2B‑Base**, a compact yet powerful vision+language model by Hugging Face. It’s designed for **multimodal understanding**—including images, multi‑image inputs, and videos—while offering **faster and more efficient inference** thanks to quantization. Perfect for on-device and resource-constrained deployments. |
|
|
|
|
|
--- |
|
|
|
|
|
### 🔧 Base Model Summary |
|
|
|
|
|
* **Name**: SmolVLM2‑2.2B‑Base |
|
|
* **Publisher**: Hugging Face TB |
|
|
* **Architecture**: Idefics3 vision encoder + SmolLM2‑1.7B text decoder |
|
|
* **Modalities**: image, multi-image, video, text |
|
|
* **Capabilities**: captioning, VQA, video analysis, diagram understanding, text-in-image reading |
|
|
|
|
|
--- |
|
|
|
|
|
### 📏 Quantization Details |
|
|
|
|
|
|
|
|
**Method**: torchao quantization |
|
|
|
|
|
**Weight Precision**: int8 |
|
|
|
|
|
**Activation Precision**: int8 dynamic |
|
|
|
|
|
**Technique**: Symmetric mapping |
|
|
|
|
|
Impact: Significant reduction in model size with minimal loss in reasoning, coding, and general instruction-following capabilities. |
|
|
--- |
|
|
|
|
|
### 🎯 Intended Use |
|
|
|
|
|
* On-device or low-VRAM systems (edge, mobile, small GPUs) |
|
|
* Multimodal tasks: VQA, captioning, comparing images, video transcription |
|
|
* Research on quantized multimodal models |
|
|
|
|
|
--- |
|
|
|
|
|
### ⚠️ Limitations & Considerations |
|
|
|
|
|
* May underperform compared to full-precision version |
|
|
* Only supports the modalities supported by the base model |
|
|
--- |
|
|
|
