Step 1: Run on Your Laptop

Get a first structured decision running on your own machine. This is the crawl stage of From Laptop to Production: ideal for writing policies, trying templates on private data, and working offline.

1. Start an engine

Pick the option that matches your hardware.

Option A: Mac with Apple Silicon (Metal), no Docker or cloud

make setup       # check prerequisites
make download    # 4-bit model pack, ~18.8 GB
make local-up    # diffgemma engine on :8080 + dgem gateway and Decision Studio on :8090
make local-status

A 32 GB Mac works with the recommended context size; memory budgeting, the engine’s own API and troubleshooting are in the Apple Silicon engine reference. Stop everything with make local-down.

Option B: Linux or Windows workstation with an NVIDIA GPU (Docker)

The serving image needs a GPU that can run the NVFP4 checkpoint (Blackwell, e.g. RTX PRO 6000, is what we test; an L4 works but is slower; see GPU requirements). There are no prebuilt images yet: first build the serving image into your own Artifact Registry (GCP_PROJECT=<PROJECT> make cloudrun-build), then run it:

gcloud auth configure-docker us-central1-docker.pkg.dev

# Lean image: downloads the public weights from Hugging Face on first boot (needs internet)
docker run --gpus all -p 8080:8080 \
  us-central1-docker.pkg.dev/<PROJECT>/dgem/dgemma:<tag>

# Or mount a local copy of the weights instead of downloading them
docker run --gpus all -p 8080:8080 -v /data/dgemma:/weights -e MODEL=/weights \
  us-central1-docker.pkg.dev/<PROJECT>/dgem/dgemma:<tag>

Wait until curl -s localhost:8080/health shows "vllm_ready": true (and "warmed": true a few seconds later).

2. Build the CLI

make build        # -> ./bin/dgem

3. Make your first decision

./bin/dgem decide \
  -t templates/support_triage.json.tmpl \
  -v 'ticket=Urgent: production database connection pool exhausted after deploy' \
  --stats

Example output (illustrative, not a recorded run):

SLOT             | TYPE       | VALUE                | PROB       | ENTROPY     | MARGIN
-----------------------------------------------------------------------------------------
sentiment        | score      | frustrated           | 99.8%      | 0.002 nats  | 1.00
team             | choice     | engineering          | 100.0%     | 0.000 nats  | 1.00
urgent           | boolean    | yes                  | 99.9%      | 0.001 nats  | 1.00

All three questions are answered in one forward pass, each with its own probabilities. support_triage opts in to 4 samples (one parallel batch) so it can report agreement; most templates use one. How to read probabilities and hesitation: Your first decision policy.

4. Open Decision Studio

make local-up already started it; otherwise run ./bin/dgem serve --port 8090 and open http://localhost:8090. Browse the built-in policies, run one on your own text or image, and inspect the per-question probabilities and the trace waterfall.

5. Connect an AI assistant (MCP)

{
  "mcpServers": {
    "dgem": { "command": "/path/to/dgem/bin/dgem", "args": ["mcp", "--local"] }
  }
}

--local targets the engine on your machine only (no cloud credentials or fallback), so start the engine first. opencode setup, other MCP clients, and troubleshooting: Studio, MCP and HTTP API.

Next

• Write your own policy: Your first decision policy.
• Need more speed or a shared endpoint? Use a remote GPU or deploy your own on Cloud Run.