The Archive · Reference

Models & providers

Local or cloud, on each surface — the engines, the providers, and where your keys are kept.

Checked on Desktop 0.1.15CLI 0.1.0Android 0.1.38

A newer version is out (Desktop 0.1.25, CLI 0.5.2, Android 0.1.40): some details may differ.

  • Local: the model runs on your own hardware, and the conversation stays on the device.
  • Cloud: the model runs at the provider you choose, with your own API key. The conversation goes straight to that provider, with no TALOS server in between.

Downloading a model or searching the web uses the network for that action only.

  • Local engine. llama.cpp is included, in a CPU and a Vulkan build. The app picks the graphics card when Vulkan lists one, otherwise the processor; the Local engine menu (Automatic · Graphics card · Processor) lets you choose.
  • Models. GGUF models download inside the app, in the Model Lab, their size and SHA-256 checked before they are used; a file that does not match is rejected.
  • Cloud. 23 providers with your own key, from OpenAI and Anthropic to Amazon Bedrock and Google Vertex AI; Ollama and LM Studio on your computer too.
The model sheet, Local tab: models that run on this computer, offline.
The model sheet, Local tab: models that run on this computer, offline.TALOS Desktop 0.1.15 · demo session
  • Local engines, with no key: Ollama, LM Studio and llama.cpp.
  • Cloud, 22 providers: OpenAI, Anthropic, Google Gemini, DeepSeek, OpenRouter, Mistral, Groq, xAI, Kimi, Qwen, Z.AI, MiniMax, Cerebras, Together, Fireworks, DeepInfra, Novita, Nebius, Hugging Face, Azure AI Foundry, Amazon Bedrock and Google Vertex AI.
  • TALOS never falls back from a local model to a cloud one on its own.
  • /provider sets up a provider, /model (or Alt+P) picks the model, and talos doctor says what is ready.
  • On the phone: GGUF models through llama.cpp inside the app. The GPU backend is offered only after the device passes a real qualification check. Before you download, each quantization is measured against the device’s memory.
  • Cloud: OpenAI, Anthropic, Gemini, DeepSeek and OpenRouter with your own key, or your own Ollama endpoint.
  • Codice, the coding agent, needs a cloud provider for now: a model on the phone cannot drive it yet.
Surface Where the key is kept
Desktop The operating system’s keyring, never the browser
CLI The operating system’s credential store, typed in hidden input; never a file. A key already in your environment is used only after you agree once.
Android With the app’s settings, on the phone: there is no TALOS backend.

Type to search the guides.