🦴 Sentinel Tiny Image — Sentinel-INT4
Production-ready quantized model for edge deployment.
This is the Sentinel-INT4 quantized variant of Sentinel Tiny Image.
📊 Quantization Spec (SDM-v1 §4)
| Property | Value |
|---|---|
| Format | Sentinel-INT4 |
| Zero-Point (Z) | C₁ = -0.007994021806 (attracting fixed point) |
| Scale (S) | max|w| · (1/e) |
| Compression | 8× |
| Size | ~1.4 MB |
| Parent model | 5dimension/sentinel-tiny-image |
🧬 Mathematical Foundation
The Sentinel quantization uses the attracting fixed point C₁ as the zero-point, minimizing rounding noise by clustering weights around the dynamical attractor.
Quantization: q = round((w - C₁) / S)
Dequantization: w = q · S + C₁
Where S = max|w| · (1/e)
🚀 Deployment Targets
| Platform | Compatible |
|---|---|
| Mobile (iOS/Android) | ✅ |
| IoT devices | ✅ |
| Microcontrollers (ARM Cortex-M) | ✅ |
| Browser (ONNX/WebGL) | ✅ |
| Edge TPU | ⚠️ (needs conversion) |
📦 Download
from huggingface_hub import hf_hub_download
# Download quantized weights
weights = hf_hub_download(repo_id="5dimension/sentinel-tiny-image-int4", filename="model_int4.pt")
🔗 Links
- Parent model: 5dimension/sentinel-tiny-image
- Main repo: sentinel-manifold-discoveries
- Sentinel Hub: Interactive demo
License: MIT | Edge-ready. Theorem-backed. 🦴
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support