Browse the HuggingFace Hub. Download with a click. Run locally: detection, segmentation, classification, depth, OCR, SAM, VLMs, ASR, TTS, diffusion. No cloud, no upload, your data never leaves your GPU.
A desktop app (Windows, macOS, Linux) for running open-source ML models locally:
| Modality | Models |
|---|---|
| Vision | DETR, YOLOS, OWLv2, SegFormer, SAM 2/3, ViT, CLIP, SigLIP |
| VLM / Multimodal | Qwen-VL, LLaVA, Moondream, Florence-2, PaliGemma |
| Speech | Whisper, MMS-TTS, SpeechT5, Bark |
| Text | Llama, Mistral, Qwen, Phi, Gemma, DeepSeek |
| Diffusion | SD 1.5/XL, FLUX, SDXL-Turbo, Kandinsky |
| Depth / OCR / DocQA | DPT, Depth Anything, TrOCR, LayoutLMv3 |
200+ model families supported via the underlying transformers + diffusers runtime.
- Visit localml.tech and download the installer for your OS.
- Run the installer.
- First launch only: the app downloads a portable Python runtime (~30 MB) and installs PyTorch + transformers (~2–5 GB depending on CPU/GPU pick). Wait ~5 minutes.
- Browse the Hub, install a model, run.
Because the app isn't yet code-signed, Windows shows a "Windows protected your PC" warning on first launch. Click More info → Run anyway. The warning won't reappear after the first install. Code signing is on the roadmap.
macOS may refuse to open the app on first run. Right-click the app in Applications → Open → confirm. After the first run, Gatekeeper remembers and you can open it normally. Apple Developer ID signing is also on the roadmap.
The .AppImage is portable. chmod +x it and double-click, or run from the terminal.
Open a new issue and pick the Bug Report template. Include:
- Your OS + version
- LocalML version (titlebar shows it)
- Steps to reproduce
- Any console output if you can grab it
New issue → Feature Request template.
The LocalML binary distribution is MIT licensed. See LICENSE. Short version: free to use, modify, redistribute. No warranties.
LocalML is in active development. Expect rough edges; please file issues when you find them.
