Experimental
MoE Foundry splits one big model into focused specialists
A Mixture of Experts model contains many expert networks. MoE Foundry measures which experts a workload actually uses, keeps a selected group per layer, and exports a smaller model to test — with a CLI, an MCP server and a published validation receipt.
GitHub activity: · Published:
What it is
Think of one large team becoming several focused teams. Each smaller team still needs the shared machinery — embeddings, attention, normalisation, tokenizer, output head — because one expert alone is not a language model.
What changed
The publication commit ships the Foundry CLI, a seven tool local MCP server, a generated MetaHarness harness with a Codex host adapter, and a usage guide. Publication checks are recorded: 37 JavaScript tests, 26 Python tests, six harness tests and 13 CLI and MCP acceptance checks, with a validation receipt listing remaining qualification gates.
Get started
Python for the engine:
git clone https://github.com/ruvnet/MoE-Foundry.git
cd MoE-Foundry
python3 -m venv .venv && source .venv/bin/activate
python -m pip install -r engine/requirements.lock
python -m pip install --no-deps -e engineExpected result: the engine installed from a lock file. Then configure your MCP client with the local stdio server documented in the usage guide.
The MCP server is local stdio only, with no shell, export, GPU or network execution tool. Inspecting config metadata does not validate tensor files.
Use it today
Practical case: a narrow workload paying for a general model. Input is your traffic; the workflow is measure expert usage, keep the used group, export and test; the output is a smaller model plus an honest retention measurement.
Acceptance test: run the fixture, then evaluate the exported specialist on the same fixture. If retention is poor, that workload needs more experts than you kept.
Push it further
Experimental commentary. Per request specialist routing is the ambitious version: a fleet of small models with a router in front. The limitation is stated in the repository — a dense model without MoE experts needs a different approach entirely.
Falsifiable test: route a mixed workload through specialists and compare end to end quality against the original model. If it drops, your routing boundary is in the wrong place.
Publish MoE Foundry with CLI, MCP, MetaHarness and usage guide