Edge Studio

Edge Studio is AtomGradient's local workbench for building and testing on-device AI products on Apple Silicon. It ships as a Python package with a local Studio UI, an API server, and CLI workflows for model preparation, local chat, and the Neural Imprint learning demo.

Start with the developer documentation when you are building an app or integrating the SDK:

• Developer site: https://atomgradient.github.io/edge-developers/

Install

Use Python 3.11 in an isolated environment before installing Edge Studio.

With venv:

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install --upgrade --pre edge-studio

With uv:

uv venv --python 3.11 .venv
source .venv/bin/activate
uv pip install --upgrade --pre edge-studio

Current release candidate: 0.0.1rc23.

For a deterministic RC install:

python -m pip install edge-studio==0.0.1rc23

The package installs one command-line entry point:

Quick Start

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade --pre edge-studio
edge studio
open http://127.0.0.1:18842

Edge Studio binds to 127.0.0.1:18842 by default. Override the host or port with VLM_HOST and VLM_PORT when you need a different local address.

First Model

The baseline developer-preview model is qwen3.5-9b-4bit.

edge models where qwen3.5-9b-4bit
edge models fetch qwen3.5-9b-4bit --source auto

edge models fetch supports ModelScope, Hugging Face, and hf-mirror sources. The auto mode chooses a source order from the local network environment.

First Chat

Run a local multi-turn chat after the model is available:

edge demo chat --model qwen3.5-9b-4bit --interactive

The first 9B model load can take tens of seconds. After [chat:ready], ask a few questions and the answer streams token by token. The CLI uses model-aware generation defaults; override the output length with --max-tokens when needed. Exit with /exit.

Learning Demo

Run the local correction-learning demo once the baseline model is ready:

edge demo learn run \
  --sample finance_conservative_cashflow_v1 \
  --model qwen3.5-9b-4bit \
  --max-tokens 160 \
  --include-text

The demo uses a synthetic finance preference to exercise local receipts, correction capture, Neural Imprint artifact generation, and a follow-up query without bundling internal evaluation data.

Export a Scaffold App

After the model is available, export an iOS scaffold app ZIP directly from the CLI. This path is designed for Code Agents and does not require clicking the Web UI:

edge export scaffold \
  --model qwen3.5-9b-4bit \
  --app-name FinanceAgent \
  --output ./exports

The command writes a stable ZIP path under ./exports. Add --bundle-id and --team-id when you want the generated Xcode project to be closer to real-device signing on the first open.

For the full walkthrough, SDK integration paths, and API references, use the Edge Developers documentation:

• https://atomgradient.github.io/edge-developers/

License

Edge Studio is distributed under the AtomGradient Proprietary License. See LICENSE.