lattice-ai/phi-4-mini-private

Lattice Phi-4 Mini Private

Privacy Tier: wrapped | Parameters: 3.8B | Context: 131,072 tokens | VRAM: ~6GB

Microsoft's newest small model — Phi-4 Mini. Punches well above its weight class in reasoning and code. MIT licensed, runs on laptops. The best edge model for private fine-tuning when GPU memory is limited.

Privacy Guarantees

Feature	Status
Sandboxed training (no network egress)	Yes
PII output guardrails	Yes
Encrypted training logs	Yes
Zero telemetry	Yes
DP-SGD training support	Yes
Privacy certificate on export	Yes

Quick Start

pip install ltce
ltce pull lattice-ai/phi-4-mini-private
ltce train ./your-data --model phi-4-mini-private --epsilon 4.8 --method qlora
ltce verify ./output/adapter

from ltce import Lattice

lt = Lattice()
vault = lt.encrypt("./sensitive-data/", password="...")
result = lt.train(
    model="phi-4-mini-private",
    data=vault,
    epsilon=4.8,
    method="qlora",
)
lt.verify(result)

What is Lattice?

Lattice is a privacy-first model training platform. The value isn't running locally (anyone can do that). The value is:

• DP-SGD training -- individual training examples can't be extracted from weights
• Signed certificates -- BLAKE3 hash + ed25519 signature proves the privacy guarantee
• Safe sharing -- publish your adapter knowing the training data is mathematically protected

Capabilities

general, reasoning, code, instruct

Base Model

microsoft/phi-4-mini-instruct

License

MIT

---

Built with Lattice -- Train private. Prove it. Share safely.