System requirements

There are two distinct system-requirements questions, and they have very different answers.

The first: what do I need to run Ertas? Ertas is a browser app; GPUs are provisioned in the cloud. The bar is low.

The second: what do my users need to run the model I trained? This is where most of the real constraints live. The Ship section lays out the specifics per platform; this page collects them into one lookup.

Running Ertas (browser-side)

Ertas runs entirely in a modern browser. There is no native install, no GPU required on your machine, and no Python.

Training itself happens on Ertas-managed T4 or A10G GPUs; nothing about your hardware affects training speed. Dataset uploads are bounded by your network upload bandwidth and your plan's dataset storage cap, both of which live on the Pricing page.

Running an Ertas-exported model on-device

This is the part where the specific hardware floors matter. The exported GGUF is what runs on the user's device, and the floor depends on the model size and the platform.

iOS

A 1B-class Q4_K_M model holds about 700 MB of native RAM plus the KV cache. A 3B-class model holds about 2.0 GB. On a 4 GB device the OS will kill the app under memory pressure; on a 6 GB device a 1B-class model is usable but a 3B-class will be tight.

For the integration mechanics, see Ship: iOS. For the delivery story (App Store size limits, on-demand resources, direct download), see Ship: model delivery and UX.

Android

Apple Intelligence requires iPhone 15 Pro (8 GB) and Pixel 8 Pro requires 12 GB for Gemini Nano. Ertas's 6 GB floor for a 1B-class model is on the lower end of the industry's positioning; surface the requirement in your store listing so users on 4 GB devices are not surprised.

Desktop

Desktop is the most forgiving target. The Ollama bundle Ertas ships works on Intel Macs, Apple Silicon Macs, Windows, and Linux out of the box. CUDA acceleration on a discrete NVIDIA GPU gives a 2-to-10x speedup over CPU; without one, modern CPUs still hit 5 to 15 tok/s on a 3B-class model.

Web (browser)

WebGPU support is the gating factor. Without it, wllama falls back to pure WebAssembly which is roughly 5x slower. Firefox enabled WebGPU support in 141 (2026); Safari shipped it in 18 / iOS 17. Older browsers still load the model but inference is uncomfortably slow.

Browser storage quotas vary. Chrome lets you use up to 80% of free disk space; Firefox is similar; Safari historically capped at 1 GB but has lifted to roughly 60% on macOS 14 and iOS 17. The 3 GB minimum reflects what a 3B-class Q4_K_M model needs plus working space; smaller models drop the requirement proportionally.

Storage planning

The model file size is the dominant storage cost, not the runtime. A summary:

The KV cache scales with sequence length: about 50 to 200 MB for 1B-class models at a typical chat context, more for larger models or longer contexts. See Performance tips for the trade-offs around context length and the KV cache size.

What's next