phi-3-mini-private

Lattice Phi-3 Mini Private

Privacy Tier: wrapped | Parameters: 3.8B | Context: 4,096 tokens | VRAM: ~6GB

Microsoft Phi-3 Mini (3.8B) — surprisingly capable small model wrapped with Lattice privacy. Runs on laptops and edge devices. Perfect for private on-device inference where VRAM is limited.

Privacy Guarantees

Quick Start

pip install ltce
ltce pull lattice-ai/phi-3-mini-private
ltce serve lattice-ai/phi-3-mini-private --port 8080 --sandbox

from ltce import Lattice

lt = Lattice()
lt.serve("lattice-ai/phi-3-mini-private", port=8080, sandbox=True)

The model serves an OpenAI-compatible API at http://localhost:8080/v1/chat/completions.

What Does "Wrapped" Mean?

This model uses the original microsoft/Phi-3-mini-4k-instruct weights — no additional training has been done.

What Lattice adds is the privacy serve layer:

• 🔒 Sandbox mode — all outbound network connections are blocked during inference
• 🛡️ PII guardrails — model output is scanned for emails, phone numbers, SSNs, credit card numbers, API keys
• 📝 Encrypted logs — inference logs are AES-256-GCM encrypted at rest
• 🚫 Zero telemetry — HuggingFace telemetry, W&B, MLflow all disabled

For models with mathematical privacy guarantees (DP-SGD training), see our hardened tier models.

Capabilities

general, reasoning, instruct

Base Model

microsoft/Phi-3-mini-4k-instruct

License

MIT

---

Built with Lattice — privacy-first local AI.