L-AI 1.20: f32 KV cache, per-model Precise kernels (auto-fix for models that only output end-of-text), polished memory box
Browse files- .gitattributes +1 -0
- L-AI-1.19.apk → L-AI-1.20.apk +1 -1
- README.md +114 -113
.gitattributes
CHANGED
|
@@ -47,3 +47,4 @@ L-AI-1.16.apk filter=lfs diff=lfs merge=lfs -text
|
|
| 47 |
L-AI-1.17.apk filter=lfs diff=lfs merge=lfs -text
|
| 48 |
L-AI-1.18.apk filter=lfs diff=lfs merge=lfs -text
|
| 49 |
L-AI-1.19.apk filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 47 |
L-AI-1.17.apk filter=lfs diff=lfs merge=lfs -text
|
| 48 |
L-AI-1.18.apk filter=lfs diff=lfs merge=lfs -text
|
| 49 |
L-AI-1.19.apk filter=lfs diff=lfs merge=lfs -text
|
| 50 |
+
L-AI-1.20.apk filter=lfs diff=lfs merge=lfs -text
|
L-AI-1.19.apk → L-AI-1.20.apk
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 65040134
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cf92ca585acdc4d5a9740b5d2a9f036c852773ab26fcf311ebd6a3eb14169475
|
| 3 |
size 65040134
|
README.md
CHANGED
|
@@ -1,113 +1,114 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: other
|
| 3 |
-
license_name: l-ai-freeware
|
| 4 |
-
tags:
|
| 5 |
-
- android
|
| 6 |
-
- apk
|
| 7 |
-
- gguf
|
| 8 |
-
- llama.cpp
|
| 9 |
-
- on-device
|
| 10 |
-
- local-llm
|
| 11 |
-
- offline
|
| 12 |
-
---
|
| 13 |
-
|
| 14 |
-
# L-AI — private AI chat that runs on your phone
|
| 15 |
-
|
| 16 |
-
**L-AI** is a free Android app by **UltraLabs** for chatting with AI models that run **fully on your device**.
|
| 17 |
-
No account and no cloud. Your chats never leave your phone unless you turn on a feature that needs the internet, such as web search.
|
| 18 |
-
It runs any **GGUF** model through a fresh build of llama.cpp. It also runs many older GGML `.bin` models that other apps no longer load.
|
| 19 |
-
|
| 20 |
-
📥 **Download:** [`L-AI-1.
|
| 21 |
-
|
| 22 |
-
---
|
| 23 |
-
|
| 24 |
-
## Highlights
|
| 25 |
-
|
| 26 |
-
- 🧠 **Any GGUF model.** Qwen, Llama, Gemma, Phi, LFM, Mistral, MoE models and more. Each model is formatted with its **own chat template**, so it behaves the way its makers intended.
|
| 27 |
-
- 🗂️ **Older models too.** Legacy GGML `.bin` files are detected and loaded automatically: LLaMA (GGJT/GGMF/ggml), GPT-2, and GPT-NeoX (StableCode, StableLM-alpha, Pythia, Dolly, RedPajama).
|
| 28 |
-
- 👁️ **Vision.** Attach photos from the gallery or camera and let vision models (with an `mmproj` file) describe them. Downloading a model from Hugging Face grabs its projector automatically (Q8, else F16). Keep several projectors: each model is linked to the one that actually fits it (checked from the files), and nothing loads while Vision is off.
|
| 29 |
-
- 💡 **Thinking models.** A 💡 button turns reasoning on or off. It only appears for models that support it. A thinking budget cuts long thoughts short, and the reasoning is shown in a collapsible "Thought process". Qwen3.5 reasoning models are detected from their metadata and their <think> block is opened for them, the way they expect.
|
| 30 |
-
- ⚡ **Speed tools.**
|
| 31 |
-
- Speculative decoding with a draft model.
|
| 32 |
-
- MoE expert offloading and prefetching.
|
| 33 |
-
- KV cache quantization (f16 / q8_0 / q4_0) and flash attention.
|
| 34 |
-
-
|
| 35 |
-
-
|
| 36 |
-
-
|
| 37 |
-
|
| 38 |
-
-
|
| 39 |
-
-
|
| 40 |
-
-
|
| 41 |
-
-
|
| 42 |
-
-
|
| 43 |
-
-
|
| 44 |
-
|
| 45 |
-
-
|
| 46 |
-
-
|
| 47 |
-
-
|
| 48 |
-
-
|
| 49 |
-
-
|
| 50 |
-
-
|
| 51 |
-
|
| 52 |
-
-
|
| 53 |
-
-
|
| 54 |
-
|
| 55 |
-
-
|
| 56 |
-
-
|
| 57 |
-
-
|
| 58 |
-
|
| 59 |
-
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
- **
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
|
| 73 |
-
|
|
| 74 |
-
|
|
| 75 |
-
|
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
- [GRAFT-1B
|
| 80 |
-
- [
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
|
| 86 |
-
|
|
| 87 |
-
|
|
| 88 |
-
|
|
| 89 |
-
|
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
|
| 97 |
-
|
|
| 98 |
-
|
|
| 99 |
-
|
|
| 100 |
-
|
|
| 101 |
-
|
|
| 102 |
-
|
|
| 103 |
-
|
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
- [
|
| 109 |
-
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
license_name: l-ai-freeware
|
| 4 |
+
tags:
|
| 5 |
+
- android
|
| 6 |
+
- apk
|
| 7 |
+
- gguf
|
| 8 |
+
- llama.cpp
|
| 9 |
+
- on-device
|
| 10 |
+
- local-llm
|
| 11 |
+
- offline
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# L-AI — private AI chat that runs on your phone
|
| 15 |
+
|
| 16 |
+
**L-AI** is a free Android app by **UltraLabs** for chatting with AI models that run **fully on your device**.
|
| 17 |
+
No account and no cloud. Your chats never leave your phone unless you turn on a feature that needs the internet, such as web search.
|
| 18 |
+
It runs any **GGUF** model through a fresh build of llama.cpp. It also runs many older GGML `.bin` models that other apps no longer load.
|
| 19 |
+
|
| 20 |
+
📥 **Download:** [`L-AI-1.20.apk`](./L-AI-1.20.apk) (65 MB) · Android 11+ · 64-bit ARM (arm64-v8a)
|
| 21 |
+
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
## Highlights
|
| 25 |
+
|
| 26 |
+
- 🧠 **Any GGUF model.** Qwen, Llama, Gemma, Phi, LFM, Mistral, MoE models and more. Each model is formatted with its **own chat template**, so it behaves the way its makers intended.
|
| 27 |
+
- 🗂️ **Older models too.** Legacy GGML `.bin` files are detected and loaded automatically: LLaMA (GGJT/GGMF/ggml), GPT-2, and GPT-NeoX (StableCode, StableLM-alpha, Pythia, Dolly, RedPajama).
|
| 28 |
+
- 👁️ **Vision.** Attach photos from the gallery or camera and let vision models (with an `mmproj` file) describe them. Downloading a model from Hugging Face grabs its projector automatically (Q8, else F16). Keep several projectors: each model is linked to the one that actually fits it (checked from the files), and nothing loads while Vision is off.
|
| 29 |
+
- 💡 **Thinking models.** A 💡 button turns reasoning on or off. It only appears for models that support it. A thinking budget cuts long thoughts short, and the reasoning is shown in a collapsible "Thought process". Qwen3.5 reasoning models are detected from their metadata and their <think> block is opened for them, the way they expect.
|
| 30 |
+
- ⚡ **Speed tools.**
|
| 31 |
+
- Speculative decoding with a draft model.
|
| 32 |
+
- MoE expert offloading and prefetching.
|
| 33 |
+
- KV cache quantization (f32 / f16 / q8_0 / q4_0) and flash attention. f32 is exact, for models that break at lower precision.
|
| 34 |
+
- **Precise kernels** per model (Auto / On / Off): if a model only answers with an instant end of text, L-AI reloads it without the fast ARM kernels, checks again, and remembers the fix only when it helps.
|
| 35 |
+
- Memory-mapped or in-RAM loading.
|
| 36 |
+
- A built-in benchmark (prompt processing and generation).
|
| 37 |
+
- 🔧 **Tools the model can call** (native tool calling):
|
| 38 |
+
- Web search (DuckDuckGo / Wikipedia).
|
| 39 |
+
- A calculator that gives exact results.
|
| 40 |
+
- File creation.
|
| 41 |
+
- 📚 **Knowledge.** Add TXT / PDF / EPUB files. The most relevant passages are found and given to the model.
|
| 42 |
+
- 🧷 **Memory.** The model can remember facts you tell it across chats. You can add, view or delete them anytime.
|
| 43 |
+
- 🎭 **Personas and slash commands.** `/summarize`, `/translate`, `/eli5`, `/fix` and your own custom ones.
|
| 44 |
+
- 💬 **A chat experience you'd expect:**
|
| 45 |
+
- Markdown, code blocks with copy and **Run** (live HTML/JS preview), charts, Mermaid diagrams and LaTeX.
|
| 46 |
+
- Regenerate with versions, edit and resend, fork a chat from any message.
|
| 47 |
+
- Read aloud, voice input with optional auto-send, and haptic ticks while it writes.
|
| 48 |
+
- ✨ **Instant follow-up suggestions** after each reply. They're worked out on the device without using the model, so they cost nothing.
|
| 49 |
+
- 🏷️ **Automatic chat names** from a tiny bundled titling model (about 20 ms, no setup).
|
| 50 |
+
- 🕶️ **Temporary chats** that are never saved or remembered.
|
| 51 |
+
- 🗃️ **History:**
|
| 52 |
+
- Search, folders, pins, share and export (.md / .txt).
|
| 53 |
+
- Backup and restore.
|
| 54 |
+
- 🔌 **Integrations:**
|
| 55 |
+
- An **OpenAI-compatible API server** on your local network, so you can use your phone's model from a PC.
|
| 56 |
+
- A **remote server** option: chat with a bigger model running on your PC (llama.cpp, LM Studio, Ollama, Jan, Atomic Chat or any OpenAI-compatible server). With Tailscale it works from anywhere, and replies stream in live.
|
| 57 |
+
- Share-to-L-AI, a Quick Settings tile and a home-screen widget you can resize from a tiny icon up to a big panel.
|
| 58 |
+
- 🎨 **Make it yours:**
|
| 59 |
+
- Material You or custom colors, 9 presets and AMOLED black.
|
| 60 |
+
- Bubble / card / minimal chat styles, fonts, text size and chat wallpapers.
|
| 61 |
+
|
| 62 |
+
## Getting started
|
| 63 |
+
|
| 64 |
+
1. Download **`L-AI-1.20.apk`** above, open it, and allow installing from your browser or file manager if Android asks.
|
| 65 |
+
2. Open L-AI, tap the model library, then either:
|
| 66 |
+
- **Download from Hugging Face** inside the app (search any GGUF repo), or
|
| 67 |
+
- **Add a model file** you already have (`.gguf` or `.bin`).
|
| 68 |
+
3. Load it and start chatting. That's it.
|
| 69 |
+
|
| 70 |
+
**Which model should I try?** On most phones, a 0.5B–4B model at Q4/Q8 is the sweet spot.
|
| 71 |
+
|
| 72 |
+
| Phone RAM | Good model sizes |
|
| 73 |
+
|---|---|
|
| 74 |
+
| 4 GB | up to ~1.5B (Q4) |
|
| 75 |
+
| 6–8 GB | up to ~4B (Q4) |
|
| 76 |
+
| 12 GB+ | 7–8B (Q4), or small MoE models |
|
| 77 |
+
|
| 78 |
+
Small models from the same creator work well:
|
| 79 |
+
- [GRAFT-1B](https://huggingface.co/SmallAICreator/GRAFT-1B)
|
| 80 |
+
- [GRAFT-1B-Vision-GGUF](https://huggingface.co/SmallAICreator/GRAFT-1B-Vision-GGUF) (with image support)
|
| 81 |
+
- [MiniGPT2-22M-GGUF](https://huggingface.co/SmallAICreator/MiniGPT2-22M-GGUF) (tiny, for fun)
|
| 82 |
+
|
| 83 |
+
## Permissions, explained
|
| 84 |
+
|
| 85 |
+
| Permission | Why |
|
| 86 |
+
|---|---|
|
| 87 |
+
| Internet / network state | Only for features you turn on: Hugging Face downloads, web search, the local API server |
|
| 88 |
+
| All files access (optional) | Load a huge model straight from storage without copying it, when your phone is low on space |
|
| 89 |
+
| Foreground service + notifications | Keep the local API server running while the app is in the background |
|
| 90 |
+
| Vibrate | Haptic ticks while the model writes (adjustable or off) |
|
| 91 |
+
|
| 92 |
+
There are no ads, no analytics and no accounts.
|
| 93 |
+
|
| 94 |
+
## Details
|
| 95 |
+
|
| 96 |
+
| | |
|
| 97 |
+
|---|---|
|
| 98 |
+
| Package | `com.ultralabs.lai` |
|
| 99 |
+
| Version | 1.20 (build 21) |
|
| 100 |
+
| Min Android | 11 (API 30) |
|
| 101 |
+
| ABI | arm64-v8a |
|
| 102 |
+
| Signing certificate | `CN=UltraLabs, O=UltraLabs` |
|
| 103 |
+
| Cert SHA-256 | `517602ea12fa156ab2713e6ca298fc856e95454ff2132c6a5267d6dec16af8c9` |
|
| 104 |
+
| APK SHA-256 | `cf92ca585acdc4d5a9740b5d2a9f036c852773ab26fcf311ebd6a3eb14169475` |
|
| 105 |
+
|
| 106 |
+
## Credits
|
| 107 |
+
|
| 108 |
+
- [llama.cpp](https://github.com/ggml-org/llama.cpp) (MIT): the inference engine.
|
| 109 |
+
- [Cyronius/titler](https://huggingface.co/Cyronius/titler) (Apache-2.0): the tiny bundled chat-naming model.
|
| 110 |
+
- Made by **SmallAICreator** · **UltraLabs**.
|
| 111 |
+
|
| 112 |
+
## License
|
| 113 |
+
|
| 114 |
+
L-AI is free to download, use and share in unmodified form. Models you load are covered by their own licenses.
|