Private LLMs

What are Private LLMs

Private LLMs are supported via a set of OpenAI-compatible APIs that run AI models within a trusted execution environment (TEE). This makes it possible to build new private AI applications or to migrate existing ones to run in a secure environment where your data remains private.

How Private LLMs Work

1. Send your prompt over HTTPS (just like any secure web request)
2. Your prompt runs inside a Trusted Execution Environment where it is hidden from the cloud provider
3. Get your result back over HTTPS, complete with a cryptographic signature

Available Models

Model	Parameters	Features	Best For	Access
meta-llama/Llama-3.2-3B-Instruct	3B	Chat completion, tool support	Quick responses, simple Q&A, basic chat. Faster inference with lower resource usage.	Testnet
meta-llama/Llama-3.1-8B-Instruct	8B	Chat completion, tool support	More complex tasks, detailed responses, code generation. Higher quality but slower inference.	Testnet
deepseek-ai/DeepSeek-R1-Distill-Qwen-14B	14B	Chat completion	Reasoning and complex tasks that require a long output. Slowest of the three.	Apply for access
hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4	70B	Chat completion, tool support	Very high-quality responses, complex reasoning, detailed generation (e.g., coding, creative writing)	Apply for access

For most use cases, start with the 3B model and move up if you need more capability.

Key Features

• Drop-in Privacy: Keep your OpenAI-style code, just point it at a private LLM
• Privacy Through TEEs: All processing happens within a TEE, built on NVIDIA Confidential Computing
• Standard Security: HTTPS encryption provides end-to-end protection for all data in transit between your application and the TEE environment
• Cryptographic Attestation: Verify the integrity of your TEE environment through our attestation API
• Signed Responses: Every model response includes a cryptographic signature for verification

Attestation & Verification

Two layers of cryptographic proof are provided:

1. Environment Attestation: Verify your TEE's integrity through the attestation endpoint, including:

   • GPU environment verification
   • Service verification through a verifying key

2. Response Signatures: Every chat completion response includes a cryptographic signature, allowing you to verify that the response came from the attested environment.

Getting Started

1. Get access to private LLMs by obtaining your nilAI API key and node URL
2. Use your API key to check the /models endpoint for available models
3. Start running private AI using your chosen model
4. Optionally, verify your environment using the attestation API

Enhance by Providing Context from Private Storage and nilRAG

Optionally, you can use nilRAG to allow private LLMs to access nilDB private storage in order to retrieve relevant context! See nilRAG documentation for details.