Run 35B MoE models in ~2.5 GB of RAM. Weights stream from SSD; a trained router predicts loads a step ahead. Mac, iPhone, Android, Windows supported
Python
NOTICE — edge0 / windows (release tree)
========================================
This file records third-party material contained in or referenced by this subproject.
Licensing of the distribution is governed by the LICENSE file at the monorepo root.
1. ggml-org/llama.cpp (MIT) — https://github.com/ggml-org/llama.cpp
The engine is the pristine upstream checkout pinned at commit 7ab4ee7 (tag b11100),
located at the monorepo root vendor/llama.cpp. All modifications ship as replayable
patch bands under patches/llama.cpp/ (never committed into the vendor tree).
2. Rust crates (see app/src-tauri/Cargo.lock for exact versions/licenses):
tauri 2, serde, reqwest (rustls), sha2, hex, windows-sys — each under its own license.
3. JS/npm packages (see app/package.json + package-lock.json): React, Vite, Tailwind,
streamdown/markdown rendering stack — each under its own license.
4. Model weights (downloaded at runtime, not part of this repository):
Edge0/Edge0-8B-A1B-preview and Edge0/Edge0-35B-A3B-preview — published in MLX format,
converted on-device to GGUF Q4_1 by the tools/ converter chain. Model-card license
applies to the weights and applies to the conversion outputs unchanged.