Welcome to Ertas
What Ertas is, who it's for, and how to navigate these docs.
Ertas is a fine-tuning platform for on-device language models. Its visual training surface (the canvas, the Run panel, and project management) is Model Studio, referred to as Studio in the rest of these docs. You bring a dataset, pick a base model, configure a training run on the Studio canvas, and walk away with a quantized GGUF file you can ship inside an iOS, Android, desktop, or web app. No GPU rental contract. No notebook setup. No DevOps team required.
These docs are written for two readers. If you are an app builder, they walk you from a blank canvas to a working model in one sitting. If you are evaluating Ertas before committing to a fine-tune run, every section doubles as a reference for what the platform actually does, with no marketing fluff. Either way, you should be able to find what you need in under a minute.
What you can build with Ertas
Anything that calls for a small, specialised, private model running on a user's device:
Customer support agents
Fine-tune a 3B model on your past tickets and ship it inside your support app.
Document assistants
Train a summariser or extractor that runs offline on a laptop or phone.
Code completions
Adapt a base model to your codebase, ship it as a desktop sidecar.
Voice and transcript cleanup
Post-process speech-to-text output without sending audio off device.
The common thread: the model is small enough to fit on-device (1B to 14B parameters, 4-bit quantised), specialised enough to be useful in its niche, and yours to deploy however you want once it leaves Ertas.
The Ertas workflow
Every project in Ertas follows the same four-stage path:
Prepare a dataset
Upload JSONL, import from Hugging Face, or generate synthetic data with your own AI tool through Prompt Studio. Datasets are validated for shape and rights attestation before training can start.
Configure a run on the canvas
Open Studio, drop a Fine-Tune or Train module on the canvas, and connect four legs: a base model, a training dataset, a training config, and a LoRA config. Each leg has sensible defaults so you can press play in under two minutes.
Train on a managed GPU
Hit play. Ertas queues your job on a T4 or A10G GPU, streams logs to the Run panel, and reports loss, throughput, and estimated cost as it goes. You only spend credits while the GPU is actually attached.
Export and ship
A successful run produces a LoRA adapter and, by default, a quantised GGUF file ready for Ollama, llama.cpp, or your platform of choice. Download it and embed it in your app.
Start with your coding agent
Prefer to build from your editor? Point your coding agent at Ertas and let it help you plan the model, the dataset, and the integration. This works in Cursor, Claude Code, Codex, OpenCode, or any coding agent. Paste the prompt below to get going.
How to navigate these docs
The sidebar mirrors the workflow above:
Get Started
Run your first fine-tune in 15 minutes.
Datasets
Formats, quality, sourcing, and synthesis.
Studio
The canvas, the run lifecycle, and project management.
Export
GGUF, quantization, and verifying your build.
Ship
Deploy to iOS, Android, desktop, and web.
Cookbook
End-to-end recipes for common use cases.
The right rail on every page is a live table of contents. The search bar (press Ctrl + K or ⌘ + K) indexes the entire docs site.
A few things Ertas is not
So you do not waste time finding out the hard way:
- Not a hosting platform. Ertas trains and exports. Inference happens on your device, your server, or wherever you choose to load the GGUF.
- Not a frontier-model trainer. Ertas is tuned for small open-weights models (1B to 14B) where LoRA adapters and consumer-tier GPUs make sense. You will not pretrain a 70B base here.
- Not a labelling tool. You bring a dataset, or you generate one with the synthetic data tools. Ertas does not crowd-source labels.
If you are still deciding whether Ertas is the right fit, the FAQ covers the most common pre-purchase questions. If you already know you want to fine-tune something, skip to the Quickstart.