01
Initialize
Start the real workload, build its runtime state, and validate the response before capture.
vLLM or SGLang
DGX Spark results are in
Bring GPU inference online without the cold start.
ColdSnap captures initialized vLLM and SGLang workloads, then restores them from a validated state—so you don’t rebuild the same model, kernels, and CUDA graphs every time.
AGPL-3.0vLLM + SGLangPre-release
ColdSnap TTFT33.1s
85.6% less wait
✓capsule verified
sha256:9f3…✓distributed state restored
2 / 2✓model hydrated
ready✓validation response passed
200Measured, not modeled
Accepted August 2026 qualification matrix
6.9×faster in the strongest measured case
36 / 36restore samples passed validation
2inference engines qualified
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.
02
Capture
Package driver-qualified process and GPU state in OCI capsules while keeping large model weights separate.
portable contracts
03
Restore
Reconstruct a compatible serving workload, hydrate its weights, validate it, and return it to your manager.
ready to serve
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 architecture01
Placement managerSparkrun today · Kubernetes-ready boundary
strict request02
ColdSnap controllerengine dispatch · compatibility · receipts
activation runtime03
GPU launch unitsCRIU · CUDA · NCCL · model payloads
Latest qualification
Less waiting. Real responses.
Time to first non-empty streamed token on two DGX Spark GB10 hosts, with Linux page cache cleared before every sample. Watch the measured medians race, then choose a workload to inspect the matrix.
time-lapse running
Same model. Same hardware. Very different wait.
6.9× faster · 16× time-lapse
Vanilla start
$ sparkrun run qwen-vllm.yaml
▋
ColdSnap · n610
$ sparkrun run qwen-vllm.yaml
▋
ColdSnap first token33.1s
ready while vanilla still starts196.8s saved
Vanilla first token229.9s
Terminal phases are a time-compressed illustration of each startup path; finish times are the measured three-sample medians. Preparation and verification before container start are excluded from both.
Explore the matrix
ColdSnap’s n610 native path was fastest in every qualified workload below.
Fastest validated path6.9×faster to first token
Qwen3.8 27B FP8 · vLLM · n610 native restore
0s115s230s
Vanilla229.9s
ColdSnap33.1s
196.8 seconds savedmedian of 3 accepted runs
Evidence before adjectives.
ColdSnap is pre-release research software. Compatibility claims are intentionally limited to qualified engine, model, CUDA, driver, hardware, and topology combinations—and published alongside the result.
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.
Capture expensive initialization once, move between active and sleeping states, and bring validated models back without starting at zero.
- DGX Spark GB10 qualification
- Sparkrun reference workflow
- Transparent results and boundaries
Hosters & GPU operators
Treat warm state as an infrastructure artifact.
Integrate strict requests, content-addressed payloads, OCI capsules, compatibility admission, and stable receipts into your own placement and lifecycle systems.
- Prepare-before-evict workflow
- Transport-neutral host-provider boundary
- Machine-readable capabilities and receipts
EnginesvLLM · SGLang
Snapshot driversn580 · n610
PackagingOCI capsules
Weight pathsNative · recovery
Start with the real thing
Two commands. One captured workload.
Sparkrun is the reference manager today. It builds the request from your recipe and placement, then coordinates capture and restore through ColdSnap’s public contracts.
# Capture initialized state once
$ sparkrun coldsnap capture recipe.yaml \
--cluster two-node
# Future runs use the ColdSnap strategy
$ sparkrun run recipe.yaml \
--cluster two-nodeOpen 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.
Licensed underAGPLversion 3 only
Created togetherIndependent engineering.
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.