# 📵 Free Liquid LFM3 — tiny AI that runs on your phone, offline

**URL:** <https://onehack.st/t/free-liquid-lfm3-tiny-ai-that-runs-on-your-phone-offline/327431>\
**Category:** Tools & Scripts\
**Tags:** programming, privacy, android\
**Created:** [November 22, 2026, 1:15pm UTC](https://onehack.st/t/free-liquid-lfm3-tiny-ai-that-runs-on-your-phone-offline/327431 "2026-11-22T13:15:31Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![Hannah](https://onehack.st/user_avatar/onehack.st/hannah/32/174875_2.png) [@Hannah](https://onehack.st/u/Hannah)\
**Post date:** [November 22, 2026, 1:15pm UTC](https://onehack.st/t/free-liquid-lfm3-tiny-ai-that-runs-on-your-phone-offline/327431/1 "2026-11-22T13:15:31Z")

</div>

📵 A whole AI brain fits in about one app of space — and it keeps thinking when your signal bars vanish.

Liquid AI’s LFM3 is free, open, and runs right on your phone or laptop. It sees photos, hears you, talks back, and can learn a job you give it.

**Pick a door — all free:**

| Door | What you get |
| --- | --- |
| 🌐 **Just peek** | [Omni in a browser tab](https://huggingface.co/spaces/LiquidAI/LFM3-Omni-2B-WebGPU) — sees, hears, speaks, no install |
| 💻 **Five minutes** | `ollama run lfm3`, then switch off the wifi and keep chatting |
| 📱 **Older phone** | [CPU-only browser build](https://huggingface.co/spaces/fusorlabs/liquidai-lmf3-wasm) — no WebGPU needed |

**Your mischief list**

- 🧠 [Make it your own expert on Google’s free GPU](https://github.com/Liquid4All/cookbook/blob/main/finetuning/notebooks/grpo_with_unsloth.ipynb) — a few hours, one job
- 📦 Grab the [1.05 GB smart 4-bit file](https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF/blob/main/LFM3-1.7B-QAD-Q4_0.gguf) — trained to keep its smarts at tiny size
- 🎙 [Talk over it and cut in](https://github.com/iamdgarcia/lfm3-omni-voice-agent) — a voice agent that lives in your browser
- ⚡ [Let it decide instead of chat](https://huggingface.co/LiquidAI/d1-2B) — many questions answered at once, with confidence scores
- 🧪 Try the [community remix](https://huggingface.co/aceflooder/LFM3-1.7B-OT3-2M-GGUF) — same brain, 1.58× faster on one Mac

> **🗝️ More free doors**
>
> - 🆓 [A 262K-memory brain by API](https://openrouter.ai/liquid/lfm-3-1.7b:free) — $0 in and out, no download
> - 🧩 [A 45-line offline chat app](https://debbie.codes/blog/offline-ai-study-app-local-models-lfm3) — no server, account or key
> - 👀 [Run the vision model on a free GPU](https://huggingface.co/LiquidAI/LFM3-VL-2B/discussions/1) · [a coding workspace in a tab](https://huggingface.co/spaces/webml-community/LFM3-1.7B-WebGPU-Pi)
> - 🤖 [A phone app that does the chores](https://github.com/makefunstuff/offline-android) — notes, to-dos, calendar, alarms, camera questions, all on-device
> - 🖥 [The full 256K memory in ~4.3 GB](https://www.toolmintx.in/blog/lfm3-1-7b-vram-256k-context-hermes-agent) · [all four models under 5 GB](https://sotaaz.com/post/lfm3-8gb-en)
> - 🎥 [Voice + webcam agent](https://github.com/varadganjoo/lfm3-omni-realtime-agent) · [a Pi making image decisions](https://agihunt.info/en/p/1a1a413cea502fc848e98bc10f4)

> **📊 The score, straight from the maker**
>
> - **71.4** — the vision model’s average over **62** vision tests with thinking on, against **70.0** for Gemini 3.5 Flash Lite
> - **62.1** — the same model with thinking off; Gemma 4 12B (71.7) and GPT-5.6 Luna (73.4) still lead the table
> - **~1 GB** is the weight file; running memory on an iPhone peaked at **1.9 GB** with a 32K prompt
> - Every figure above is Liquid’s own; [the announcement](https://www.liquid.ai/blog/lfm3) and [the scoreboard written up](https://xiaohu.ai/en/article/liquid-lfm3/) carry them
> - Free commercial use runs to **$10M** a year in revenue; [the weights](https://huggingface.co/collections/LiquidAI/lfm3) are open

Go on — switch off the wifi.
