GitHub radar
35B model in 3GB RAM with Edge0 on your Mac
Announcing Edge0’s 35B sparse MoE model, which streams experts from SSD and uses only 3GB of active memory on Apple Silicon.
We have trained a 35B sparse MoE model (Qwen3.5-35B) that can run on Apple Silicon with SSD offloading of experts. This means the model weights are stored on disk and loaded into memory when needed (only 3GB of active memory), rather than loading all the weights into RAM at once like a normal 35B model that would require 64GB+ of RAM. Our entire model weights take up 19.6GB of disk space. Our model can generate text at ~15 tokens per second and read inputs (prompt completion) at ~140 tokens per second. You can interact with our model using a custom CLI interface that you can easily install with pip. You can also use our model in both chat mode and API server mode.
Why a vibe-coder should care
Now you can run a 35B scale model locally on your Apple Silicon Mac with just ~20GB of SSD space! Previously you had to have 64GB of RAM or access to a GPU server to run such large models locally. Our model performs similarly to other open 35B models available online, which should suffice for playing around, simple repetitive tasks, and understanding what these open models can do. For serious work, we recommend using Claude or Fable instead.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install the Edge0 model locally: https://huggingface.co/Edge0/Edge0-35B-A3B-preview — set up the edge0 CLI package per the README, launch chat mode and show me how to interact with it.
Requires an Apple Silicon Mac with ~20 GB of free SSD space — the model stores weights on disk and loads the needed parts into RAM during generation, total RAM usage is around 3 GB.
Open on Hugging Face▌ More finds