The September 4 v0.3.18 matrix completed functional qualification across Qwen3.8 27B FP8 on vLLM and SGLang and DeepSeek V4 Flash 0731 on vLLM. It exercised vanilla startup plus n580 and n610 recovery and native restores.
Every primary matrix sample returned the exact expected validation response. Ten of the twelve ColdSnap restore performance rows passed their comparison threshold; two DeepSeek native rows remain performance follow-ups.
Why functional and performance status are separate
A restore can be correct without meeting the current performance guardrail. Keeping those outcomes separate prevents a performance regression from being mislabeled as a compatibility failure, while still making it visible before a release is accepted.
The DeepSeek n580 native path was effectively tied with recovery in this run. DeepSeek n610 native remained clearly faster than recovery, but was slower than its comparable retained-graph baseline. Both cases are under investigation.
Portability evidence
The qualification also moved Qwen/vLLM n580 capsules between n580 capture hosts and different n610 hosts without recapture. The manager materialized target-local state and verified the restored inference response on the destination placement.
That evidence covers a specific declared route; it does not make every capsule universally portable. Host capabilities, snapshot driver selection, native payload admission, runtime images, and topology remain part of the contract.
See the benchmark hub for the accepted public headline and the link to the complete September qualification report.
Browse the ColdSnap documentation or inspect the source on GitHub.