Ship on-device AI in your app. No cloud API required.
Train a custom model, export it to GGUF, and run it on your users' devices. No per-token bill, no cloud latency, no user data leaving the phone.
Free plan to start · Paid plans from $10/mo when you scale · Cancel anytime
The Problem
Cloud AI Doesn't Scale in Mobile Apps
Every API call is a cost, a latency hit, and a privacy risk. Here's what that means as your app grows.
API costs scale with your users
At 1,000 users your OpenAI bill is $120/month. At 10K it's $620. At 40K it's $3,000+. Every new user is a line item, and margins vanish the moment your app gains traction.
Cloud latency kills your UX
Every AI feature in your app adds a network round-trip. Users see loading spinners, dropped requests on poor connections, and complete failure when offline.
User data leaves the device
Every cloud API call sends your users' input to a third-party server. That's a privacy risk, a compliance headache, and a trust gap your competitors can exploit.
One deprecation away from broken
When OpenAI retires a model, your app breaks on their timeline. Rate limits throttle you at scale. You're building on rented infrastructure with no fallback.
How It Works
From Dataset to On-Device in 4 Steps
Cloud GPUs handle the training. Your users' phones handle the inference.
Upload your dataset
Drag in your training data — conversation logs, product docs, domain examples. Start with as few as 100 high-quality samples.
Train on our GPUs
Pick a mobile-sized base model, configure the training recipe visually, and fine-tune on cloud GPUs.
Export as GGUF
One-click export to GGUF format. The open standard supported by llama.cpp, Ollama, and every major local inference runtime.
Run on-device
Bundle the GGUF model in your iOS or Android app. llama.cpp handles inference natively. Zero per-token cost, forever.
Entrena tu modelo de IA. Visualmente.
What You Get
Built for App Builders, Not ML Engineers
The full pipeline from training data to on-device model: no code, no ML expertise, no stitching tools together.
Visual training canvas
Upload your dataset, pick a base model, configure the recipe, and fine-tune without Python, YAML, or Jupyter notebooks.
One-click GGUF export
Export your fine-tuned model as GGUF — the open format compatible with llama.cpp, Ollama, and LM Studio. No proprietary lock-in.
Deploy on iOS and Android
GGUF models run natively on mobile devices via llama.cpp. Your users get instant inference with zero network dependency.
Privacy by architecture
Train in the cloud, run on-device. User data never leaves the phone at inference time. Ship with "your data stays on your device" as a real feature, not marketing.
Domain-specific accuracy
A fine-tuned 3-7B model beats GPT-4 prompting on your specific task. 94% accuracy on domain tasks vs 71% with prompt engineering alone.
Works offline, zero latency
No network required at inference time. On-device means instant responses — even in airplane mode, underground, or on spotty connections.
The Math
Ertas + On-Device vs Cloud API
Side by side at 10,000 monthly active users.
| Feature | Ertas (On-Device) | Cloud API |
|---|---|---|
| Cost at 10,000 users | $25/mo (flat) | $620+/mo (scaling) |
| Per-inference cost | $0 | $0.0075-0.03/1K tokens |
| Latency | Instant (on-device) | 200-800ms (network) |
| Offline capability | Full | None |
| User data sent to third party | No | Yes |
| Model ownership | You own the GGUF file | Rented access |
| Deprecation risk | None — you have the file | High |
"Fine-tuned a model on our product docs in under an hour. Now our support bot actually understands our domain instead of hallucinating."
Jamie K.
Indie Developer
"Replaced our $400/mo API bill with a fine-tuned model running locally. Better results, predictable costs. Exactly what we needed."
Maria R.
Startup Founder
Ship AI that runs on your users' devices
Ship a real on-device model on the free plan. Upgrade to a paid plan from $10/mo to train bigger models on more capable GPUs. No per-inference cost, ever.