Self-host on Lambda

Clean per-hour GPU instances (A10 → H100 → B200) with simple pricing and fast provisioning. You get a raw Ubuntu box: SSH in, run vLLM.

Credentials

cloud.lambda.ai → API keys → LAMBDA_API_KEY (for their instance API) — the console alone is enough for manual use. Add an SSH key under SSH keys.

Deploy

Console → Instances → Launch — pick a GPU (1× A10 runs 8-9B; 1× H100 runs 32B), your SSH key, launch
SSH in and serve:

ssh ubuntu@<instance-ip>
pip install vllm
vllm serve Qwen/Qwen3-32B --host 0.0.0.0 --port 8000 \
  --enable-auto-tool-choice --tool-call-parser hermes

Reach it — pick one:
- SSH tunnel (simplest, secure): ssh -N -L 8000:localhost:8000 ubuntu@<ip> then /connect http://localhost:8000/v1 Qwen/Qwen3-32B
- Open the port in the instance firewall and connect directly (put --api-key sk-something on vLLM if you do)

Teardown

Terminate the instance in the console — billing is per-hour while it exists, idle or not. Weights re-download next time (or park them on a persistent filesystem).

Let your agent do it

mkdir -p ~/.claude/skills/selfhost-lambda
curl -o ~/.claude/skills/selfhost-lambda/SKILL.md https://mantisagent.cc/skills/lambda.md