Quickstart

Go from local code to running on GPUs in under two minutes.

Prerequisites

Make sure you've completed Installation — you'll need a Mithril API key, project and billing configured before continuing.

1. Start from your project

Open a terminal in the directory that contains your project code:

cd my-ml-project

2. Describe your task in YAML

Create a file called task.yaml at the project root. This tells Mithril what hardware you need, how to prepare the environment, and what to run.

Tip: Run ml launch with no arguments to generate a starter task.yaml and AGENTS.md file that coding agents can use to author workload specs and automate setup.

# task.yaml

# Sync this directory so your code and data are
# available on the cluster
workdir: .

resources:
  infra: mithril
  accelerators: B200:8
  
# Maximum hourly price you're willing to pay for
# the instance.
# Due to auction-based pricing, you often pay less
# than this cap.
config:
  mithril:
    # Equivalent to $4.00/GPU/hour on an 8x instance.
    limit_price: 32.00

# Runs once when the cluster is first created
# (install deps, download data, etc.)
setup: |
  pip install -r requirements.txt

# Command that executes your code — runs on the
# cluster every time you launch or exec.
run: |
  nvidia-smi
  python train.py

This example shows training, but the same workflow applies to batch inference and evaluation.

That's the entire configuration. workdir: . uploads your local files, setup runs once to prepare the environment, and run specifies the commands that executes your code.

→ Task YAML — full reference manual.

3. Launch

ml launch task.yaml -c my-cluster

This does two things: creates a cluster named my-cluster with the GPUs you requested, and submits your task as the first job. Logs will stream directly to your terminal.

4. Check cluster status

Once launched, you can check on your cluster at any time:

ml status

5. View the job queue and logs

A cluster can run multiple jobs. To see what's queued and stream output:

ml queue my-cluster        # list all jobs on the cluster
ml logs my-cluster         # stream logs for a job

6. SSH into the cluster

For interactive debugging or poking around, SSH straight in.

ssh my-cluster

7. Iterate on your code

Edit your code or run commands locally, then push the changes to your existing cluster:

ml exec my-cluster task.yaml

ml exec syncs your workdir and executes the run commands without reprovisioning or modifying cluster resources. This creates a tight iteration loop for testing changes.

When to use which:

Scenario	Command
Only your code or run commands changed	ml exec my-cluster task.yaml
You changed setup, file_mounts, or resources	ml launch task.yaml -c my-cluster

8. Tear down when you're done

When you no longer need the cluster:

ml down my-cluster

This terminates your bid, releasing compute resources and stopping billing. Persistent volumes are not affected — your data remains available for future clusters.