Cribl cuts privacy-model latency 63%

Cribl says its Guard Privacy Model now runs at 2.8× throughput with 63% lower median latency after pruning 20 of 32 attention heads, while F1 moved only 0.06 percentage points in development benchmarks. The Cribl compression report says measurements were FP32 PyTorch before production quantization and threshold tuning, so treat gains as a development signal rather than an SLA. For teams scanning telemetry or other fixed-label workloads, this is a reminder to benchmark task-specific pruning before paying for a larger general model; it complements Microsoft’s small-model security-routing case.