Overview

What you can do with it

Everyday Q&A, writing, and rewriting.

Explaining, summarizing, and multi-turn exchanges over text.

Code-related questions and reasoning tasks.

Once a model is ready, inference runs locally on the device without using a cloud model service.

Check the model’s outputModel-generated content may be inaccurate. For important facts, professional decisions, or code execution, verify the results yourself.

Two model tiers

Model	Characteristics	When to choose it
Edge0 8B	About 7.9B total parameters, about 1.2B active, based on `inclusionAI/Ling-3.0-tiny-base`	Smaller size and resource footprint; a good fit for a first try
Edge0 35B	About 35B total parameters, only a small subset active per step, based on `Qwen/Qwen3.6-35B-A3B`	Trying a larger model when the device has sufficient resources

For a full comparison, see Models: 8B vs 35B.

How local inference works

Edge0 uses a sparse mixture-of-experts (MoE) model. For each generated token, the model uses only a subset of its experts and reads them from device storage on demand, so it does not need to hold all weights in memory at once.

Each platform uses an inference stack suited to its hardware:

Platform	How it runs
macOS	MLX / Metal inference on Apple Silicon, with a companion desktop app
Windows	A local engine based on llama.cpp, using Vulkan to accelerate part of the computation
iOS	MLX Swift / Metal inference on iPhone
Android	Based on llama.cpp, running inference on the CPU

Because of this, model preparation and the entry points differ by platform. Follow the guide for your platform: macOS · Windows · iOS · Android.

Edge0

Trust & Policies

Shared

What you can do with it

Two model tiers

How local inference works

Current release

Start exploring

​What you can do with it

​Two model tiers

​How local inference works

​Current release

​Start exploring

What you can do with it

Two model tiers

How local inference works

Current release

Start exploring