GitHub radar
35B model in 3GB RAM — Edge0 runs it on your Mac
Edge0 released a 35B sparse MoE model that streams expert weights from SSD on demand, keeping active RAM use to just 3GB on Apple Silicon.
Edge0 built a 35B sparse MoE model (based on Qwen3.5-35B) that runs on Apple Silicon using SSD expert offloading: weights are stored on disk and streamed into RAM on demand, so active memory stays around 3GB. A standard 35B model would require 64GB+ of RAM to hold all weights at once. The full checkpoint is 19.6GB on disk. Generation speed is roughly 15 tokens per second for responses; prompt filling (reading the input) reaches 140 tokens per second. The model is controlled via a dedicated CLI tool installed through pip, and works in chat or API server mode.
Why a vibe-coder should care
If you have an Apple Silicon Mac with ~20GB of free SSD space, you can now run a 35B-class model locally — previously that took a 64GB RAM machine or a server GPU. Quality is what you'd expect from an open 35B model: solid for experiments, simple repetitive tasks, or when you want to understand what open models of this size can do. For real work, Claude or Fable will be faster and more reliable.
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.
Нужен Mac на Apple Silicon с ~20 ГБ свободного места на SSD — модель хранит веса на диске и загружает нужные части в оперативную память по ходу генерации, суммарная нагрузка на RAM около 3 ГБ.
Open on Hugging Face▌ More finds