Development
Delivery Models
Global 
The WhiteLightning project started with a question: Does text classification need the cloud every time?
It began as an internal experiment during our ML hack days, aimed at exploring what could be done with less. LLMs are great, but for most real-world use cases, you don’t need a 175B-parameter model on standby. Instead, you need something fast, portable, and private. Something that works offline, ships inside your app, and doesn’t rack up API bills.
Instead of utilizing LLMs at runtime, we use them once to generate synthetic data. Then, distill that into a compact, ONNX-based model that runs anywhere. No cloud, no lock-in, no friction. Just a simple way to go from idea to working classifier on your terms.
Uses LLMs once for data generation (~$0.01 vs $1–10 per query)
Easily fits in mobile apps, kiosks, or embedded firmware
Generate a binary classifier on a laptop in minutes
That’s 0.38 ms per input on commodity CPUs
Runs on low-power hardware like Raspberry Pi Zero
Identical logits across Python, Rust, Swift, and more
No cloud, no vendor lock-in, no latency risks
ONNX.js (web), iOS/Android (mobile), MCUs (embedded), laptops (desktop)
Running LLMs is expensive. They’re cloud-bound, latency-prone, and often impractical for edge or privacy-sensitive environments. Our team saw a rising demand for fast, cost-free, offline text classification across use cases from email spam filters to IVR routing and parental control on offline consoles.
Existing options couldn’t cut it:
We create a completely different tool: a command-line solution that distills LLMs into tiny, production-ready classifiers, which are fast, local, and open.
WhiteLightning isn’t a hosted SaaS. It’s a developer-first, no-nonsense CLI tool that does one thing extremely well: turn your prompt into an embeddable classifier. Here’s what’s under the hood:
These models can detect spam, classify voice commands while deploying, like config files.
It’s not a typical “tinyML” framework. WhiteLightning delivers on real-world performance:
We built WhiteLightning to be easy to use: one CLI command to get started, clean output you can actually make sense of, and workflows that plug right into your CI without hassle.
WhiteLightning is made for builders who don’t want to rent intelligence from the cloud.
WhiteLightning was developed to make intelligent text classification possible anywhere, even in environments where cloud access is limited, restricted, or simply not allowed.
Personal productivity and desktop apps
Comms safety and moderation
Healthcare and life sciences
Customer support and compliance
IoT, automotive, and smart devices
Retail and eCommerce
Developer and DevOps tools
Education
OEM/Embedded hardware
WhiteLightning was built to work even on extremely low-spec hardware. With models under 1 MB, no runtime dependencies, and ONNX compatibility, it runs smoothly on:
If it can run Python or Rust, it can run WhiteLightning.
WhiteLightning is 100% open-source under GPL-3.0. The classifiers it generates are MIT-licensed and yours to use in commercial apps.
Our team maintains it publicly:
While still early in its adoption, WhiteLightning is poised for:
We’re also experimenting with image classification, but text is where our solution shines.
