Run published capsules

For a published community capsule, materialize it locally, then run it. First complete the ColdSnap plugin preview setup, then run these commands in that activated environment.

1. Choose a qualified recipe

Search published recipes from the CLI:

sparkrun recipe search @coldsnap

You can also browse the ColdSnap recipes on GitHub. Choose a recipe for your hardware and keep its pinned model revision, runtime image, topology, and engine arguments.

This example uses Qwen 3.8 27B FP8 on vLLM, TP2, across two compatible DGX Spark GB10 hosts. Replace two-node with your saved cluster name. The recipe requires exactly two nodes.

2. Materialize locally

sparkrun coldsnap materialize --cluster two-node \
  @coldsnap/qwen3.8-27b-fp8-coldsnap-tp2-vllm

Sparkrun prepares published capsules, model data, caches, and any required local runtime state for the target placement. It selects the newest snapshot driver supported across all placed hosts: n580 for a lowest host driver major of 580–609, or n610 when all hosts support 610 or newer.

Materialization can pull large artifacts and perform a verification restore. Allow time and disk space for this preparation; its duration is separate from the startup TTFT measurement.

3. Run the prepared recipe

sparkrun run --cluster two-node \
  @coldsnap/qwen3.8-27b-fp8-coldsnap-tp2-vllm

Sparkrun selects the ColdSnap strategy and starts the compatible prepared workload. Use the serving endpoint reported by the run and confirm the expected inference response. Later starts use the same sparkrun run command; changes to your recipe or placement may require preparation again.

4. Check status or stop

sparkrun status --cluster two-node
sparkrun stop --cluster two-node \
  @coldsnap/qwen3.8-27b-fp8-coldsnap-tp2-vllm

To release GPU memory between uses or prepare a workload ahead of demand, see sleep, wake, and warm. To create a capsule for your own workload, follow capture and share.