Web-LLM Chat
hash Tools
This is a chat application that runs a Large-Scale Language Model (LLM) on the device.
- Downloads
- 42+
- on Google Play
- Price
- Free
- No in-app purchases
- Updated
- Aug 7
- 2026 · v1.1.4
Screenshots
About Web-LLM Chat
This Android app uses Web-LLM with HTML5 (WebGPU) and Capacitor.
Chat content is never sent externally. Communication only occurs during model downloads and model list updates. After model download, it can be used offline.
## Screen and Usage
The top bar always displays a RAM meter and menu.
The RAM meter shows the device's available memory (updated every 2 seconds) and indicates whether the model will fit in memory.
### Chat Tab
Converse with one model at a time. Simply select a model and send a message; if not already loaded, it will automatically load before the conversation begins.
The [Load]/[Release] buttons are for pre-loading and manual release.
### 3-Model Comparison Tab
The screen is divided into three panes, allowing you to ask a single question to up to three models simultaneously and compare their answers side-by-side. Because smartphones have only one GPU and limited memory, models are generated sequentially one at a time (other models are released just before generation, leaving only one model resident).
### Model Tab
Download/Delete/Refresh List/Filter/Generation Settings for Models.
- A maximum of 3 models can be saved on your device. Downloaded models work offline.
- [Refresh List]: Checks the distribution status of the distributor (Hugging Face) and detects models that are no longer available.
Even if distribution has ended, downloaded models can still be used on your device (they will remain in the list with a badge).
- [Filter ▾]: Narrow the list by checking FP16 / FP32 / Coder / Math.
- Badge Meaning:
- Generation Settings: Context length (default 1024, memory saving) / Maximum generated tokens (default 512) / Temperature (default 0.7). These will be applied from the next load.
### Menu
- New Chat: Resets the conversation in the currently displayed tab.
- History: List of conversation history (saved on your device). Tap to resume, [Delete] to delete.
The 3-model comparison session is also restored to the pane for each model.
- Language: Switch display language (selection is saved on the device)
- Version Info: Version / WebGPU diagnostics (adapter information, measured buffer limit) / OSS license
## Model Catalog
The list is automatically derived at startup from the built-in catalog of the included web-llm library.
Adding new models requires an app update.
- FP16: Memory-saving. However, some mobile GPU drivers have a bug in FP16 calculations that may cause garbled characters.
- FP32: Approximately 1.2 to 2 times more memory, but numerically more stable.
- Qwen3 series generates "think" tokens, but this app suppresses the generation of thought tokens.
## System Requirements
- A WebGPU-compatible WebView (Android System WebView / a newer version of Chrome) is required. A warning banner will be displayed at startup on unsupported devices.
- Model weights are stored in the WebView's Cache API (hundreds of MB to 2 GB/model). Please be mindful of your device's storage space.
## Privacy
Chat history, settings, and model data are all stored only on your device.
Chat content is never sent externally. Communication only occurs during model downloads and model list updates. After model download, it can be used offline.
## Screen and Usage
The top bar always displays a RAM meter and menu.
The RAM meter shows the device's available memory (updated every 2 seconds) and indicates whether the model will fit in memory.
### Chat Tab
Converse with one model at a time. Simply select a model and send a message; if not already loaded, it will automatically load before the conversation begins.
The [Load]/[Release] buttons are for pre-loading and manual release.
### 3-Model Comparison Tab
The screen is divided into three panes, allowing you to ask a single question to up to three models simultaneously and compare their answers side-by-side. Because smartphones have only one GPU and limited memory, models are generated sequentially one at a time (other models are released just before generation, leaving only one model resident).
### Model Tab
Download/Delete/Refresh List/Filter/Generation Settings for Models.
- A maximum of 3 models can be saved on your device. Downloaded models work offline.
- [Refresh List]: Checks the distribution status of the distributor (Hugging Face) and detects models that are no longer available.
Even if distribution has ended, downloaded models can still be used on your device (they will remain in the list with a badge).
- [Filter ▾]: Narrow the list by checking FP16 / FP32 / Coder / Math.
- Badge Meaning:
- Generation Settings: Context length (default 1024, memory saving) / Maximum generated tokens (default 512) / Temperature (default 0.7). These will be applied from the next load.
### Menu
- New Chat: Resets the conversation in the currently displayed tab.
- History: List of conversation history (saved on your device). Tap to resume, [Delete] to delete.
The 3-model comparison session is also restored to the pane for each model.
- Language: Switch display language (selection is saved on the device)
- Version Info: Version / WebGPU diagnostics (adapter information, measured buffer limit) / OSS license
## Model Catalog
The list is automatically derived at startup from the built-in catalog of the included web-llm library.
Adding new models requires an app update.
- FP16: Memory-saving. However, some mobile GPU drivers have a bug in FP16 calculations that may cause garbled characters.
- FP32: Approximately 1.2 to 2 times more memory, but numerically more stable.
- Qwen3 series generates "think" tokens, but this app suppresses the generation of thought tokens.
## System Requirements
- A WebGPU-compatible WebView (Android System WebView / a newer version of Chrome) is required. A warning banner will be displayed at startup on unsupported devices.
- Model weights are stored in the WebView's Cache API (hundreds of MB to 2 GB/model). Please be mindful of your device's storage space.
## Privacy
Chat history, settings, and model data are all stored only on your device.
What’s new
Version 1.1.4 · Updated August 7, 2026
最初のリリース。
Alternatives to Web-LLM Chat
Looking for apps like Web-LLM Chat? Here are 6 similar Android apps on Google Play. The highest rated is ISA Flashcards: Arborist Exam (4.8/5), and all 6 are free to download.
-
TicNote: AI Recorder & Agent 3.2
-
dub: Social Investing Unlocked 4.5
-
ISA Flashcards: Arborist Exam 4.8
-
GitMind: AI Mind Mapping App 4.0
-
Hedy: AI Meeting Coach & Notes 4.8
-
FAX Desk: Send Fax from Phone 4.5
| App | Rating | Price | Installs | Last updated | Compare |
|---|---|---|---|---|---|
| Web-LLM Chat | – | Free | 10+ | Aug 7, 2026 | |
| TicNote: AI Recorder & Agent | 3.2 / 5 | Free | 10,000+ | Oct 30, 2026 | Compare |
| dub: Social Investing Unlocked | 4.5 / 5 | Free | 100,000+ | Oct 23, 2026 | Compare |
| ISA Flashcards: Arborist Exam | 4.8 / 5 | Free | 1,000+ | Sep 14, 2026 | Compare |
| GitMind: AI Mind Mapping App | 4.0 / 5 | Free | 1,000,000+ | Sep 27, 2026 | Compare |
| Hedy: AI Meeting Coach & Notes | 4.8 / 5 | Free | 10,000+ | Nov 6, 2026 | Compare |
| FAX Desk: Send Fax from Phone | 4.5 / 5 | Free | 10,000+ | Dec 18, 2025 | Compare |
Compare Web-LLM Chat with the top alternatives: ratings, updates, ASO score and keyword gaps →
Similar apps on Google Play
- TicNote: AI Recorder & Agent
- dub: Social Investing Unlocked
- ISA Flashcards: Arborist Exam
- GitMind: AI Mind Mapping App
- Hedy: AI Meeting Coach & Notes
- FAX Desk: Send Fax from Phone
- Fabric: your second brain
- KultureCity
- Meeting.ai: Work Assistant
- Liner - AI Search and Finance
- Pluralsight
- OnePageCRM - Simple CRM System
- RAFA - AI Investment Agents
- myReach - AI Assistant