metadata
title: README
emoji: 🚀
colorFrom: indigo
colorTo: yellow
sdk: static
pinned: true
thumbnail: >-
https://cdn-uploads.huggingface.co/production/uploads/6634fc18d94421fe1c02f97c/48breLiEtms1xr-xl36dc.png
short_description: Embedl - efficient AI for the edge
Embedl
Embedl develops advanced tools and algorithms for Edge AI. Our mission is to make AI models run faster, more energy-efficient, and reliably across diverse hardware platforms, while significantly reducing development time.
We help teams deploy high-performance AI on real-world, resource-constrained devices.
Embedl Models (Community)
Pre-optimized models that can be used off-the-shelf or customized for specific hardware target supported by the embedl-models package.
First release highlights:
- The fastest Small Language Models (SLMs) using FlashHead, a novel architectural improvement to the language-model head
- Works with popular models like Llama, Gemma, and Qwen
- Provides speedups on top of:
- Quantization
- Flash Attention
- Other standard optimizations
Device: Nvidia Jetson Thor
| Model | Generation speed (tokens/s) |
|---|---|
| embedl/Llama-3.2-3B-Instruct-FlashHead-W4A16 | 100 |
| Llama-3.2-3B-Instruct-W4A16* | 80 |
| RedHatAI/Llama-3.2-3B-Instruct-FP8 | 64 |
| meta-llama/Llama-3.2-3B-Instruct | 37 |
*Embedl quantized model for benchmarking similar to the FlashHead-W4A16 but without the faster FlashHead and custom generation loop.
Contact
Headquarters (Sweden)
Gamla Almedalsvägen 39
412 63 Gothenburg, Sweden
Email: info@embedl.com