Hanzo Edge
Run a model on the machine in front of you.
- Zen model
- zen-eco 4B
- GGUF
- Q4_K_M about 2.5GB on disk
- It uses what it finds
- Metal Detected at startup
- The prompt never leaves
- Nothing is uploaded
On-device inference
Where it runs
What the runtime does
One Rust binary, a core crate to embed, a WASM build for the browser, and a local server for everything else
A local inference endpoint on a port you pick. Point a client at it and the rest of your code does not change.
GGUF is the native format — Q4_K, Q5_K, Q8_0. A 4B model at Q4_K_M is about 2.5GB on disk, which is what makes a phone a plausible place to run one.
Metal on Apple Silicon, CUDA where there is an NVIDIA card, AVX2 or AVX-512 on CPU. Detected at startup rather than configured by you.
zen3-nano at 600M for embedded work and zen-eco at 4B for phones and tablets, both published pre-quantized.
Streaming over server-sent events from the local server, or a callback from the Rust API, so a UI starts drawing before generation finishes.
Nothing is uploaded and there is no key to leak. The prompt and the output stay in the process that made them, which is the whole reason to run a model here.
What runs on yours
It uses what it finds
Apple Silicon
- Uses
- Metal
Metal on Apple Silicon. Detected at startup rather than configured by you.
When to reach for it
Use Edge when the data cannot leave, the network cannot be relied on, or a per-call bill is the wrong shape. Use Hanzo Engine when you need full precision or many users at once.
Apps that work with no signal
The model is a file on disk, so the feature does not stop existing when the connection does.
- A desktop app that drafts and summarizes without a round trip
- Field tools on a laptop, on a site with no coverage
- Anything that has to keep working on a plane
Text that cannot leave the building
Some prompts are a legal problem the moment they cross a network boundary. This is the answer to those.
- Clinical and legal text processed on the workstation that holds it
- On-premise deployments with no egress at all
- Personal data handled on the device it belongs to
Inside the browser tab
The WebAssembly build runs the model in the page, with nothing behind it — so a visitor costs you bandwidth once and compute never.
- A demo that does not bill per visitor
- An assistant inside a web app you already ship
- Classification done client-side, before anything is uploaded
Small hardware
ARM64 Linux is a production target. Cortex-A class boards are experimental, and the 600M model is the one that fits them.
- A 600M model on a single-board computer
- An ARM server doing batch work on its own
- Kiosks and appliances with no account to sign in to
Get Edge
Edge compute
Related capabilities
Paste and ship
Paste this into any agent. It reads the skill manifest and calls Edge from there.
Read https://hanzo.ai/skill.md and use Hanzo Edge in my project.