Bring GPU inference online without the cold start.

ColdSnap reuses captured runtime state to bring compatible vLLM and SGLang workloads online faster. Run a published capsule with Sparkrun, or capture and share your own.

Run published capsules View the source

AGPL-3.0vLLM + SGLangSparkrun integration

vLLMTP2 · 2 hostsn610 native

10.7×faster startup

91% less wait

vs default loading · time to first token

ColdSnap 51.6s

Default loading552.5s

500.9s of waiting saved.

GPU state restoreddone

distributed state restoreddone

model weights hydrateddone

first token serveddone

Real workloads.
Real speedups.

Measured on NVIDIA DGX Spark.

Respond soonerGet to your first token faster.

Reuse the workSkip repeating runtime setup.

Share fast startsPublish capsules others can run.

The idea

Pay the startup cost once.

Model construction, compilation, calibration, distributed setup, and CUDA-graph capture are valuable work. Repeating all of it for every start is not.

Tune

Refine your model, engine settings, and parallelism. Validate quality and performance to establish your gold configuration.

vLLM or SGLang

Capture

ColdSnap initializes that configuration through its capture workflow and packages reusable runtime state into a capsule.

portable contracts

Restore

Restore the capsule on compatible hosts and get back to serving faster.

ready to serve

Real performance

Less waiting. Real responses.

Watch ColdSnap get to the first response sooner on two DGX Spark hosts. Compare startup paths across models, engines, and drivers below.

scroll to play

Same model and hardware. Faster first token.

10.7× faster · 24× playback

DeepSeek V4 Flash 0731vLLM · TP22× DGX Spark GB10

ColdSnap · n610

00:00.0

Default loading

00:00.0

ColdSnap first token51.6s

ready while default loading continues500.9s saved

Default first token552.5s

Explore the matrix

Compare default loading, InstantTensor, recovery, and native on both snapshot drivers. The multiplier compares the fastest ColdSnap path with default loading.

Best path vs default10.7×faster to first token

DeepSeek V4 Flash 0731 · vLLM · n610 native restore

0s276s553s

Default loading552.5s

InstantTensor292.4s

ColdSnap recovery64.3s

ColdSnap native51.6s

500.9 seconds saved vs default

Built for the whole path

From a Spark on your desk to GPU fleets.

The first public target is practical utility on NVIDIA DGX Spark. The architecture is designed around the contracts operators need to qualify, integrate, and automate the same lifecycle across infrastructure.

Builders & researchers

Make local GPU systems feel ready, not idle.

Materialize community capsules on your hosts and run them with Sparkrun. Capture your own workloads, measure their startup performance, and share the results.

DGX Spark GB10 qualification
Sparkrun reference workflow
Transparent results and boundaries

Start with Sparkrun

Hosters & GPU operators

Get more from every GPU.

Restore workloads on compatible hosts through your existing scheduler and serving infrastructure. Put GPU capacity to work sooner without rebuilding your stack.

Faster scale-up and scale-out
Higher deployment density with less idle GPU capacity
Faster recovery after hardware failures and service disruptions

Discuss an integration

A clean systems boundary

Your orchestrator stays in charge.

ColdSnap is a manager-driven infrastructure primitive, not another cluster daemon. Your placement manager owns scheduling, credentials, transport, and workload identity. ColdSnap owns strict capture, compatibility, restore, and receipt contracts.

Read the architecture

Placement managerSparkrun today · extensible manager API

strict request

ColdSnap controllerengine dispatch · compatibility · receipts

activation runtime

GPU launch unitsCRIU · CUDA · NCCL · model payloads

Start with the real thing

Materialize locally. Then run.

Start with a published recipe and the ColdSnap plugin preview setup for Sparkrun. Materialization prepares it for your hosts; later starts use the same run command.

For your own work: capture, run and measure, then publish when you are satisfied with performance. First-time materialization is preparation time, outside startup TTFT.

Follow the community quickstart Capture and share your work

# Prepare a published recipe locally
sparkrun coldsnap materialize --cluster two-node \
  @coldsnap/qwen3.8-27b-fp8-coldsnap-tp2-vllm

# Run on your prepared placement
sparkrun run --cluster two-node \
  @coldsnap/qwen3.8-27b-fp8-coldsnap-tp2-vllm

Open source by design

Your infrastructure should stay yours.

ColdSnap is released under AGPL-3.0-only. You can inspect it, run it, change it, and build on it—and the license protects the community’s access to improvements deployed as a network service.

Explore on GitHub Read the license

Licensed underAGPLversion 3 only

Created together

Independent engineering.
Shared commitment.

ColdSnap is a joint work of Scitrera LLC andFox Engine Ltd—built with the rigor to be useful in the open and operable in demanding environments.Talk with the team.