Auto setup
The guided setup profiles RAM, CPU cores, GPU availability, and backend preference before downloading a compatible llama.cpp or Hugging Face model path. Progress is resumable after app restarts.
Setup guide - v0.3.7
Install NativeLab and start with a working model.
NativeLab can run with local GGUF models, Hugging Face Transformers, Ollama, or API providers. The app keeps setup inside the same model registry, downloader, and runtime controls used after first launch.
Install
Default install is the GUI path.
The package keeps the default download focused on the desktop app. CLI and Labs extras are available when you need them, without forcing heavyweight transformer dependencies into the default install.
pip install nativelab
nativelab
pip install "nativelab[cli]"
pip install "nativelab[labs]"
First run
Pick the setup route that matches the machine.
Manual llama.cpp
Use the Download tab to install a matching llama.cpp build, choose a GGUF model, and register it in the shared local model registry.
API or Ollama
Register an OpenAI-compatible API, Anthropic profile, custom endpoint, or an existing Ollama model when the runtime already lives outside NativeLab.
Model sources
One registry, several runtimes.
GGUF with llama.cpp
NativeLab downloads or reuses llama.cpp, stores binaries under
llama/bin/, and keeps downloaded models under
localllm/. CUDA, Vulkan, ROCm, Metal, and CPU-only
paths are selected according to detected hardware.
Hugging Face Transformers
The Hugging Face downloader supports snapshot downloads, resume controls, gated-repo tokens, and an in-app library installer for users who started with only the base NativeLab package.
Ollama and API profiles
Existing Ollama daemons and cloud providers are registered as model entries, so chat, Labs, pipelines, and CLI status use the same model selection layer.
Hardware
Start small, scale when the machine can handle it.
NativeLab favors a working baseline over oversized defaults. Context limits, GPU layers, and model size can be raised later from the same server controls when more memory is available.
Low memory
Use a small quantized GGUF model, CPU or partial GPU offload, and conservative context.
Workstation
Use larger models, more context, full GPU offload, and pipeline-heavy workflows.
Next
Build with the runtime you just configured.
Once a model is registered, the same backend can power chat, document work, Labs, CLI commands, visual pipelines, and the AI Pipeline Builder.