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
edge0 Windows app — Tauri 2 shell
The desktop shell: model download (into ~/.edge0) → on-device MLX→GGUF conversion → supervised
llama-server child process → streamed Markdown chat.
Dev run
# 0) prerequisites: node + cargo installed; engine built (see ../scripts/vendor-build.ps1)
cd app
npm install
npx tauri dev # window (vite:1420 + rust shell)
Headless end-to-end (isolated test home, re-runnable — Range-resume keeps it idempotent):
cd app\src-tauri
cargo test # fast suite: fake-source Range resume tri-state
cargo test --test real8b -- --ignored --nocapture # full chain: 4.5 GB download -> convert -> engine -> chat
Environment overrides (every path is overridable — test isolation is a first-class discipline)
| env | default | purpose |
|---|---|---|
EDGE0_HOME |
%USERPROFILE%\.edge0 |
app home (models/state/logs/tmp) |
EDGE0_REPO |
this repo checkout | where tools/ conversion scripts live |
EDGE0_PY |
python |
interpreter for the converter chain |
EDGE0_BIN_DIR |
depot engine build dir (…/wt/win/build-vk/bin/Release) |
directory containing llama-server.exe |
EDGE0_PHYS_MEM_GB |
detected (conservative) | override for the memory-tier clamp |
EDGE0_TEST_SOURCE |
modelscope |
source pin for headless tests (huggingface also supported) |
Home layout (~/.edge0)
models/edge0-<tier>/ download target (dir name is a hard contract of the converter;
per-file sha256 verified before atomic rename)
models/edge0-<tier>-gguf/ conversion output (gguf + adapter + manifest.json idempotency gate)
state/downloads/<t>.json resume checkpoints state/verified-<t>.json per-file sha ledger
state/models.json installed-model registry
logs/engine-*.log, logs/convert-*.log engine & converter output
tmp/<t>.<key>.part in-flight download parts
Proof-of-life telemetry
We require positive proof that a mechanism is actually active before trusting a number (measured culture: past incidents taught us silent no-ops look identical to success):
- after loading 35B, the engine log must show
POOL2 initwith the resolver line, or the UI marks the pool as not-in-effect; - headless tests assert three things together: phase=ready + registered artifact sha + chat returning real Markdown structure (tables/code blocks).
Packaging notes
The installer currently ships the shell only; the engine and the Python converter are expected on the build/dev machine (env-overridable paths above). Bundling engine+converter into the installer, code signing, and an auto-updater are on the roadmap.