System Requirements
Python
| Requirement | Value |
|---|---|
| Minimum version | Python 3.11 |
| Supported versions | 3.11, 3.12, 3.13 |
Operating System
The SDK is pure Python with no platform-native extensions. It runs on any OS supported by onnxruntime:
| OS | Supported |
|---|---|
| Linux (x86-64, ARM64) | ✅ |
| macOS (Intel, Apple Silicon) | ✅ |
| Windows (x86-64) | ✅ |
Hardware
| Resource | Minimum | Notes |
|---|---|---|
| CPU | Any modern x86-64 or ARM64 | No GPU required — all inference is CPU-only |
| RAM | +512 MB | Additional overhead on top of your application; see Memory below |
| Disk | 350 MB | For bundled inference models included in the package |
| Network | Optional | Required only for Mend Platform policy sync and audit events |
Memory
Memory usage depends on which guardrails are configured. Models are loaded once per process and cached for its lifetime.
| Guardrail | Approximate memory |
|---|---|
| Jailbreak + Prompt Injection (shared model) | ~300–400 MB |
| Harmful Content | ~40–55 MB |
| Harmful Content — semantic similarity (opt-in) | ~35–50 MB |
| PII | ~50–80 MB |
| Keywords / Regex / Secret Keys / URL Filter | < 5 MB |
These figures are additive overhead on top of your existing application's memory. All standard guardrails combined add ~390–540 MB to the host process RSS.
Recommended headroom: 512 MB above your baseline application footprint. Add ~50 MB if the harmful_requests semantic similarity option is enabled.
!!! tip The first request after process start pays a one-time model load cost (1–5 s). Subsequent requests are served from the in-memory cache.
Disk
The SDK package bundles all inference models. No internet access or separate download is required at runtime.
Total bundled model footprint: ~340 MB.