README / README.md
hletrd's picture
Org card: correct the GPTQ and KL claims; link the code and the v0.1.0 releases
18a01bf verified
|
Raw History Blame Contribute Delete
3.15 kB
---
title: README
emoji: 🔩
colorFrom: gray
colorTo: green
sdk: static
pinned: false
---
<div align="center">
<img class="block dark:hidden mx-auto" src="https://raw.githubusercontent.com/XylolabsInc/.github/main/profile/assets/mark_light.png" width="80" alt="Xylolabs">
<img class="hidden dark:block mx-auto" src="https://raw.githubusercontent.com/XylolabsInc/.github/main/profile/assets/mark_dark.png" width="80" alt="Xylolabs">
<h1>Xylolabs Inc.</h1>
<p><strong>Edge Physical AI that catches equipment faults and runs robots.</strong></p>
<p>Sensors, edge AI and robots for port cranes, power plants and factories.</p>
<a href="https://xylolabs.com/en/"><img src="https://img.shields.io/badge/xylolabs.com-website-0d1117?style=flat-square" alt="Website"></a>
<a href="https://github.com/XylolabsInc"><img src="https://img.shields.io/badge/GitHub-XylolabsInc-0d1117?style=flat-square&logo=github" alt="GitHub"></a>
<a href="https://xylolabs.com/en/contact/?type=poc"><img src="https://img.shields.io/badge/Request_a-PoC-0d1117?style=flat-square" alt="Request a PoC"></a>
</div>
## On Hugging Face
Open models quantized for Tenstorrent Tensix NPUs. Most store their linear layers in TT-NN's block floating point
formats, `bfloat4_b` (BFP4) or `bfloat8_b` (BFP8), most by GPTQ and some by round to nearest, stored so TT-NN
packs the weights without changing them; a few are bf16, with no quantization. Beside them are component archives
of speech, image, video and robot models, and llama.cpp GGUF files we use as references. Each model card lists the
formats per tensor class; the language model cards also give the KL and answer check against bf16, beside
llama.cpp's Q4_K_M of the same model (Q4_0 for Google's QAT checkpoints, MXFP4 for gpt-oss).
The code is on GitHub, Apache-2.0: [tensix-bfp-gptq](https://github.com/XylolabsInc/tensix-bfp-gptq) quantizes and
checks the weights; [tensix-kernels](https://github.com/XylolabsInc/tensix-kernels) and
[tensix-runner](https://github.com/XylolabsInc/tensix-runner) hold the kernels and the runner for Tensix. Releases: [tensix-bfp-gptq v0.1.0](https://github.com/XylolabsInc/tensix-bfp-gptq/releases/tag/v0.1.0),
[tensix-kernels v0.1.0](https://github.com/XylolabsInc/tensix-kernels/releases/tag/v0.1.0),
[tensix-runner v0.1.0](https://github.com/XylolabsInc/tensix-runner/releases/tag/v0.1.0).
Collections: [one per model family](https://huggingface.co/Xylolabs/collections).
## What we build
| Platform | What it does |
|---|---|
| [XyloZero](https://xylolabs.com/en/solutions/xylo-zero/) | Edge AI that reads sensor signals with PLC and drive data to catch equipment anomalies early |
| [Battery Fire Early Detection](https://xylolabs.com/en/solutions/battery-early-warning/) | Hears the acoustic precursors of thermal runaway in AGV, AMR and ESS batteries |
| [Autonomy & Robotics](https://xylolabs.com/en/solutions/robotics/) | On-device AI for robot and drone patrols, VLA manipulation learning and port crane autonomy |
Started as signal research at Seoul National University in 2019, incorporated in December 2025.
Not affiliated with or endorsed by Tenstorrent.