Container activation
The timer begins at Docker container State.StartedAt. Manager preparation, verification, and capsule pulls performed before that point are excluded.
Measured startup performance
Faster first tokens. Every loading path.
Compare default loading, InstantTensor, ColdSnap recovery, and ColdSnap native on the same qualified workloads. All times are time to first token in seconds; lower is better.
Largest benefit vs default loading10.7× fasterDeepSeek V4 Flash 0731 · vLLM · n610 native552.5s → 51.6s
Latest reported results
Both drivers. Every loading path.
Qwen3.8 27B FP8 and DeepSeek V4 Flash 0731 · two DGX Spark GB10 hosts · ARM64 · CUDA 13 · TP2.
Largest benefit vs default10.7×default TTFT ÷ fastest ColdSnap TTFT
Validation66 / 66original qualification samples
MeasurementFirst tokenpage cache cleared before each run
TTFT in seconds. DeepSeek/vLLM n610 default and native use the recorded runs; other cells use three-sample medians. Default uses the engine's normal loader, with safetensors in the recorded vLLM run. InstantTensor is a separate vLLM baseline. N/A means the pinned SGLang image does not provide that loader. Scroll the table horizontally on small screens. ColdSnap recovery loads from the original Hugging Face weights, while ColdSnap native loads from a native weights pack materialized by ColdSnap.
| Model / engine | Driver | Default | InstantTensor | Recovery | Native | Best vs default |
|---|---|---|---|---|---|---|
| Qwen 3.8 27B FP8 / vLLM | n580 | 254.5 | 165.3 | 75.9 | 74.4 | 3.4× (native) |
| Qwen 3.8 27B FP8 / vLLM | n610 | 226.7 | 231.3 | 39.5 | 37.5 | 6.0× (native) |
| DeepSeek V4 Flash 0731 / vLLM | n580 | 424.0 | 164.6 | 136.3 | 137.6 | 3.1× (recovery) |
| DeepSeek V4 Flash 0731 / vLLM | n610 | 552.5 | 292.4 | 64.3 | 51.6 | 10.7× (native) |
| Qwen 3.8 27B FP8 / SGLang | n580 | 142.7 | N/A | 116.3 | 49.8 | 2.9× (native) |
| Qwen 3.8 27B FP8 / SGLang | n610 | 138.4 | N/A | 57.7 | 40.2 | 3.4× (native) |
The headline compares with default loading. DeepSeek/vLLM on n610 takes 552.5s by default, 292.4s with InstantTensor, and 51.6s with ColdSnap native: 10.7× versus default and 5.7× versus InstantTensor. Recovery is slightly faster than native for DeepSeek/vLLM on n580; native is an option, not a requirement for benefit.
Methodology
What the timer includes.
Page cache is cleared before each run, and completed responses must pass exact validation. Preparation and first-time downloads are separate.
First streamed token
The timer ends at the first non-empty streamed model token. The completed response must also match the expected validation response for the sample to pass.
Matched conditions
Use the same model, runtime, hardware, driver, topology, and cache conditions. Clear Linux page cache on both hosts before each sample and report the median.