Local AI

Genesis can generate text, images and speech and transcribe audio on your own computer: no accounts, no internet (once the model is downloaded), no per-use cost, and your files never leave your machine.

#Getting started

1. Open Settings → Local AI. You'll see your computer's memory and, in each category, the recommended model for it.
2. Click Download. The download shows up in the jobs panel; if it's interrupted or you pause it, Resume picks up where it left off.
3. Done: if you don't have an account connected for that task, Genesis uses the local model automatically. You can also pick Local AI as the provider in the AI Studio, or a local/… model in the Assistant.

The engines (llama.cpp, stable-diffusion.cpp and whisper.cpp) come with the installer. The speech engine (Piper) is downloaded automatically the first time you use a voice.

#Models

• The Assistant with a local model uses Genesis's tools (cut video, create documents…). Qwen3 handles them best; small models make more mistakes on long tasks.
• Graphics card:
  • On Apple Silicon the engines use the GPU (Metal).
  • On Windows and Linux the installer also ships Vulkan builds of the text and image engines, which work with NVIDIA, AMD and Intel cards and their regular drivers. Genesis checks which GPU you have and uses it automatically: it's usually 3 to 10 times faster than the CPU, and FLUX becomes practical.
  • If the driver fails at startup, Genesis falls back to the CPU on its own and says so in Settings → Local AI. You can also turn off Use the graphics card there.
  • A GPU raises the recommended effort level.
• Models with vision (Vision label) understand images without internet. With them:
  • the local Assistant answers questions about images you attach; if the chosen model can't see, it automatically uses the vision one;
  • ai.describe_image describes a photo or three frames of a video and extracts the visible text;
  • the Library finds photos and videos by what they show (option Describe photos and videos with AI).
• The Search model enables search by meaning in the Library (see BIBLIOTECA.md).
• Models are stored in <Genesis data>/models and can be deleted from the same screen.

#Video and music

• Video with Wan:
  • it's generated by the same engine as images (stable-diffusion.cpp) and Genesis turns it into MP4;
  • Wan 2.1 1.3B makes 2 to 5 second clips at 480p from text;
  • Wan 2.2 5B can also animate an image (in the Gallery, Animate on an image);
  • both share the text encoder, which is only downloaded once;
  • it needs a graphics card: it works on the CPU, but a clip can take hours.
• Music with MusicGen:
  • runs inside Genesis with ONNX Runtime, which ships with the installer;
  • composes instrumental tracks up to 30 seconds long from a description;
  • works without a graphics card; the longer the track, the longer it takes;
  • understands English descriptions best ("lofi hip hop with soft piano, 80 bpm");
  • its license doesn't allow commercial use. For a commercial project, use an account (ElevenLabs Music or Stable Audio).
• Both appear in the AI Studio and in ai.generate_video / ai.generate_music (also via MCP). They're used automatically when no account is connected.

#Effort

The Local AI effort slider, in Settings → Local AI, decides how hard your computer works:

• The recommended level is preselected based on the computer's RAM and GPU:
  • less than 8 GB → Minimum;
  • 8–15 GB → Balanced;
  • 16 GB → High;
  • 16 GB with a GPU, or 32 GB or more → Maximum.
• Background tasks (indexing the Library, describing photos, summarizing) never go above Balanced, so they don't slow your computer down while you work.
• Small models answer with grammar-constrained JSON (llama.cpp response_format): they get the format right the first time and the request doesn't need repeating.

Also from the CLI:

genesis run local.set_effort level=minimum

#Your own (fine-tuned) models

You can use your own GGUF model; for example, one fine-tuned to write your scripts or describe your products. Go to Settings → Local AI → Import my model and choose what it's for: text, vision (with its mmproj) or search. For that task it's used before the catalog models. How to train one: MODELOS_PROPIOS.md.

#From the CLI and MCP

genesis run local.models

genesis run local.download id=qwen3-4b

genesis chat --model local/qwen3-4b "summarize this document: report.docx"

Any assistant connected via MCP can ask Genesis for images or speech even with no accounts: ai.generate_image, ai.speech and ai.chat fall back to the local model when there's no other provider.

#Environment variables