[L] Large Models Layer - OpenLiate

Overview

Layer [L] abstracts foundational models behind a standardized client interface (SarvamClient in src/llm/). It dynamically normalizes model identifiers and handles token limits, temperature, and multi-turn message formatting.

Supported Providers

1. Sarvam AI (Sovereign Primary)

OpenLiate natively defaults to Sarvam AI, India’s frontier sovereign AI models:

sarvam/sarvam-105b — High-capacity multi-step reasoning.
sarvam/sarvam-2b — Ultra-lightweight edge execution.
Direct integration via hub/llm/sarvam.json.

2. Ollama & Local Silicon

Run completely offline on local Apple Silicon or NVIDIA edge devices.

3. Cloud Frontier Models

Anthropic (Claude 3.5 Sonnet) and OpenAI (GPT-4o).

Configuration in `agent.json`

Layer [L] can be declared as a clean string or a detailed object:

{
  "L": "sarvam/sarvam-105b"
}

{
  "L": {
    "provider": "sarvam",
    "model": "sarvam-2b",
    "temperature": 0.2,
    "maxTokens": 4096
  }
}