Edgi: Offline AI, On-Device

21 reviews
Content rating
Everyone
1K+
Downloads
Content rating
Everyone
Learn more
Screenshot image
Screenshot image
Screenshot image
Screenshot image
Screenshot image
Screenshot image
Screenshot image

About this app

Edgi is a home for open-source AI models that run entirely on your phone, including models that have not been available offline on Android. No account, no cloud, nothing leaves your device.

We port models the rest of mobile AI ignores and make them run offline on everyday Android phones. Browse the in-app catalog, download a model once, and it's yours to use forever, in airplane mode, on a plane, anywhere.

What you can run, all on-device:

• Chat: A growing library of open language models (LLMs): SmolLM2, Granite, Mamba, OpenELM, LFM2.5, Qwen2.5. Pick a tiny fast one or a larger, smarter one to match your phone.

• Vision: Show it a photo and ask about it. On-device vision models describe images and answer questions about what they see.

• Speech to text: Transcribe audio with Whisper, fully offline.

• Text to speech: Offline text to speech with dozens of natural voices and multiple engines (Kokoro, MMS-TTS, Pocket-TTS).

• Music: An AI music generator that runs offline — create short original clips from a text prompt with MusicGen.

• Translation: Translate speech across languages with Meta's SeamlessM4T.

• Voice cloning: Capture a voice from a short clip and have it read any text, on-device (OpenVoice).

Many of these have never been free and fully offline on Android before. Edgi brings them together in one place.

Why Edgi is different:

• A real catalog. Not one locked-in model, a browsable, growing collection of open-source models, each with clear sizes so you choose what to install.

• Truly private. Your prompts, microphone audio, photos, and results are processed on your device and are never uploaded. No servers receive your content. No sign-in.

• Works offline. Once a model is downloaded, it runs with no internet.

• Built for budget phones. Edgi runs on its own custom inference engine tuned for Qualcomm Adreno GPUs tuned for budget hardware.

OPEN SOURCE
Edgi runs on adreno-llms, our own OpenCL inference engine for Qualcomm
Adreno GPUs, open source and yours to inspect or build on:
https://github.com/a8nova/adreno-llms
Run LLMs, vision, speech & music AI on-device. Private, offline, no cloud.
Updated on
Oct 31, 2026

Data safety

Safety starts with understanding how developers collect and share your data. Data privacy and security practices may vary based on your use, region, and age. The developer provided this information and may update it over time.
  • No data shared with third parties
    Learn more about how developers declare sharing
  • No data collected
    Learn more about how developers declare collection

Ratings and reviews

21 reviews
am ahm
November 4, 2026
⭐⭐⭐⭐⭐ Excellent app! Bonsai 27B runs at a very impressive speed. The developer is very responsive to user feedback and quickly fixed the maximum output token limit issue I reported. I really appreciate the developer's commitment to improving the app and listening to users. I would also like to suggest adding Bonsai 27B 2 in a future update. Highly recommended!
Did you find this helpful?
Abel Shepherd
July 28, 2026
The creator is very responsive; they fixed a bug for me within a week. Also, this is the best local AI that I have ever tested for mobile. If you need offline AI, then I highly recommend Edgi.
1 person found this review helpful
Did you find this helpful?
NNOpt
July 29, 2026
Thank you for the kind words, means a lot! 🙏 🙏
Bemnet Alemayehu
July 21, 2026
This app seems to work pretty well on my Galaxy S21+. I use it for live transcription in English!
Did you find this helpful?
NNOpt
July 29, 2026
Thanks for trying Edgi and for taking the time to review it. Means a lot!

What’s new

Longer answers: Bonsai models can now write replies of up to 8,192 tokens, and Bonsai 27B up to 16,384 on phones with enough memory. Raise Max tokens in the model's settings. Follow-up questions after a long reply now start in seconds instead of timing out, and Stop no longer reloads the model.