Skip to main content
Kyro is an open-source product fromTonmoy InfrastructureOwned byEshan Roy
Kyrov0.1.0
Get Kyro
Kyro Docs
Model & Settings Guide

Model Providers & Configuration

Kyro includes an intelligent ModelRouter with exponential backoff retry logic, multimodal payload validation, and capability mapping.

Where Credentials Are Stored

When you enter an API key in Kyro's Settings Modal, it is immediately migrated to your Operating System Keyring (Keychain on macOS, Secret Service on Linux, Credential Manager on Windows). Keys are never written in plaintext into SQLite database files or exposed to the frontend.

Default & Custom Endpoints

Kyro's default endpoint is NVIDIA NIM API, offering ultra-fast inference for Nemotron and Llama vision models:

  • • Base URL: https://integrate.api.nvidia.com/v1
  • • Default Text Model: nvidia/nemotron-3-ultra-550b-a55b
  • • Default Vision Model: meta/llama-3.2-11b-vision-instruct

You can specify any custom OpenAI-compatible endpoint (e.g. self-hosted Ollama, vLLM, or LM Studio) by setting the custom Base URL in Settings.

Verified Model Capability Registry

Model IdentifierProviderVisionTool CallingStatus
nvidia/nemotron-3-ultra-550b-a55bNVIDIANoYesVerified
meta/llama-3.2-11b-vision-instructMeta / NVIDIAYesYesVerified
meta/llama-3.2-90b-vision-instructMeta / NVIDIAYesYesVerified
openai/gpt-4oOpenAIYesYesVerified
anthropic/claude-3-5-sonnetAnthropicYesYesVerified
google/gemini-1.5-proGoogleYesYesVerified

Multimodal Image Validation

Before dispatching image attachments to models, Kyro's router inspects the raw payload:

  • • Supported MIME types: PNG, JPEG, WEBP, GIF.
  • • Size cap: Maximum 20MB per image payload.
  • • Capability check: Refuses to transmit images to models that do not declare vision capability.