What's Changed

• Temporarily remove media and storage permissions from AndroidManifest by @copybara-service[bot] in #1076
• Refactor showBenchmarkButton configuration to isBenchmarkSupported and showBenchmarkActionButton by @copybara-service[bot] in #1080
• Add support for extra data files in the Android model allowlist. This change introduces the ModelDataFile data class to represent supplementary files that need to be downloaded alongside a main model. by @copybara-service[bot] in #1088
• Automatically reset chat session with previous history when resuming after interruption by @copybara-service[bot] in #1093
• Use hugging face api to fetch model size. by @copybara-service[bot] in #1090
• Updating Terms of Service statement. by @copybara-service[bot] in #1096
• Introduce LlmSessionManager to orchestrate LLM chat sessions and persistence. by @copybara-service[bot] in #1095
• Add model allowlist for 1.0.19 by @kkGOOG in #1106
• Improve gated model permission flow and license acknowledgment on Android. by @copybara-service[bot] in #1092
• Fix stale model initialization race condition when navigating to custom tasks. by @copybara-service[bot] in #1132
• Refactor Model initialization to use StateFlow by @copybara-service[bot] in #1138
• Increment AI Edge Gallery Android versionCode to 42. by @copybara-service[bot] in #1145
• Disable learn-something-new skill by default on non-Gemma4-E4B models. by @copybara-service[bot] in #1147
• Fall back to internal CoroutineScope in AICoreModelHelper when null is provided. by @copybara-service[bot] in #1149
• Only show benchmark button when model is an LLM. by @copybara-service[bot] in #1153
• Add reusable multi-step OnboardingDialog component with persistent DataStore state. by @copybara-service[bot] in #1159
• Support task-filtered extra data files. by @copybara-service[bot] in #1160
• Increment AI Edge Gallery Android versionCode to 43. by @copybara-service[bot] in #1162
• • Replace custom getter on ChatViewModel.currentSessionId that delegated to runtimeExecutor.activeSessionId with a standard mutable property. by @copybara-service[bot] in #1164
• Update Kotlin to 2.2.21 in Gallery Android app. by @copybara-service[bot] in #1168
• Add Scrapbook custom task. by @copybara-service[bot] in #1167

Full Changelog: 1.0.18...1.0.19

• Multimodal Chat: AI Chat now supports images and audio as well!
• Hugging Face Imports: Seamlessly import LiteRT-LM models using Hugging Face model card URLs.

• Now speaking your language! We've added translation and localization support for Canada and Brazil—with more languages coming soon!
• Your feedback matters: Share your thoughts on model responses instantly using our new custom Feedback Collector.
• Bug fixes.

What's Changed

1. Add support for importing and manage litert-lm models from Hugging face.

• users are now able to import any litert-lm model from hugging face and manage the usage of them with in AI Edge Gallery

Full Changelog: https://github.com/google-ai-edge/gallery/commits/1.0.16

What's Changed

• Dynamically adjust default system prompt based on MCP tools in Agent Skills. The agent skill task now uses the default system prompt tailored for "skills only" if no MCP tools are loaded. This helps address some skill loading issue. by @copybara-service[bot] in #880

Full Changelog: 1.0.14...1.0.15

What's Changed

1. Experimental Model Context Protocol (MCP) Support

• Introduced experimental support for MCP.
• Added a user permission flow for MCP tool calls, ensuring users are prompted for approval before the agent executes a tool (with an option to "Always allow").
• Added comprehensive documentation for MCP.

2. Hardware & Performance Enhancements

• Pixel TPU Support: Enabled execution support for models on Pixel TPUs, including support for sideloaded models.
• Speculative Decoding: Added configuration options and engine initialization for speculative decoding to improve model generation speed.

3. New Agent Skills & Capabilities

• Calendar Integration: Added new skills to create-calendar-event and read-calendar-events directly from the chat.
• Scheduled Notifications: Added a schedule-notification skill (for one-time or daily alerts), complete with a Notification Management Screen and deep-link support that opens the agent chat with a pre-filled query when tapped.
• Learn Something New: Introduced a new learn-something-new skill.
• Note: Several older skills (like calculate-hash, text-spinner, send-email) were disabled by default to refine the default agent experience.

4. UI & Chat Experience Improvements

• Gemini-like UI: Updated the UI for both Chat and Prompt Lab to better match the official Gemini app experience.
• System Prompt Customization: Added UI integration and core storage for users to edit and retrieve dynamic System Prompts.
• Chat History: Introduced chat history saving that supports text, images, and audio messages.
• Media Handling: Added a new feature allowing users to download and share images directly from the chat.

5. Notable Bug Fixes & Stability

• Fixed session reset issues when turning off skills or deleting chat history.
• Switched from using "exact alarms" to "inexact alarms" to improve Android permission compliance.
• Fixed a NumberFormatException crash in the Benchmark Results Viewer that occurred for non-US locales (e.g., parsing "8186,03").

New Contributors

• @kuishou68 made their first contribution in #629
• @yijie-yang made their first contribution in #837

Full Changelog: 1.0.13...1.0.14

• Support Gemma 4's built-in Multi Token Prediction (MTP) to accelerate decoding. Check out the blog post for more details. Try it in AI Chat, Ask Image, Ask Audio, and Prompt Lab (turn on the "Support speculative decoding" in the config dialog).
• Bug fixes.

• Add Gemma3 1B NPU support for Qualcomm Technologies SoCs. Try it in Prompt Lab! Check out the blog post for more info.

  If your device is supported, you will see a new "NPU" option in PromptLab -> Gemma3-1B-IT model card:
  

  • Installing via Play Store? No action needed; the correct libraries are bundled with the app automatically for supported SoCs.

  • Installing via APK here? Check the supported device list below for each file to ensure compatibility with your device.

    ai-edge-gallery-sm8550.apk:

    • Samsung: Galaxy S23, S23+, S23 Ultra, Galaxy Z Fold 5, Galaxy Z Flip 5, Galaxy Tab S9 Series.
    • OnePlus: OnePlus 11, OnePlus 12R.
    • Xiaomi: Xiaomi 13, 13 Pro, 13 Ultra.
    • Motorola: Edge 40 Pro (Edge+ 2023), Razr 40 Ultra (Razr+).
    • Gaming/Niche: ASUS ROG Phone 7, RedMagic 8 Pro, Sony Xperia 1 V, iQOO 11.

    ai-edge-gallery-sm8650.apk:

    • Samsung: Galaxy S24 Ultra (globally), Galaxy S24/S24+ (US/Canada/China versions), Galaxy Z Fold 6, Galaxy Z Flip 6.
    • OnePlus: OnePlus 12.
    • Xiaomi: Xiaomi 14, 14 Pro, 14 Ultra.
    • ASUS: Zenfone 11 Ultra, ROG Phone 8 / 8 Pro.
    • Others: RedMagic 9 Pro, Honor Magic 6 Pro, iQOO 12, Nubia Z60 Ultra.

    ai-edge-gallery-sm8750.apk:

    • Samsung: Galaxy S25 Ultra (and S25/S25+ in most regions).
    • Xiaomi: Xiaomi 15, 15 Pro.
    • OnePlus: OnePlus 13.
    • Realme: GT 7 Pro.
    • ASUS: ROG Phone 9 Series.
    • Others: RedMagic 10 Pro, iQOO 13, Honor Magic 7 Series.

    ai-edge-gallery-sm8850.apk:

    • Samsung: Galaxy S26, S26+, and S26 Ultra (Snapdragon version used globally for Ultra; regional for S26/S26+).
    • OnePlus: OnePlus Ace 6T (released late 2025), OnePlus 15R.
    • Vivo: Vivo S50 Pro Mini, Vivo X300 FE.
    • iQOO: iQOO 15R, iQOO Z11 Turbo.
    • Honor: Honor GT 2 Pro.
    • Motorola: Motorola Signature, Motorola Razr Fold (2026 models).

• Introducing Gemma 4: Experience the latest high-performance models running fully offline.
• Agent Skills: Extend LLMs with modular tools like display interactive maps and search Wikipedia. Supporting custom skill loading from the community. Learn more about how to create skills, and share with the community
• Thinking Mode in AI Chat: Visualize the model’s reasoning process for deeper transparency. (Note: Currently exclusive to supported models, including the Gemma 4 family).

• A new fine-tuned Mobile Actions Function Gemma 270M model to try.
• A faster and more reliable fine-tuned Tiny Garden Function Gemma 270M model to try.
• A dedicated benchmark screen to run model benchmarks directly in-app, located in the Models Management screen via the top-left menu.
• Streamlined Gemma model downloads — no HuggingFace login required.

Full Changelog: 1.0.9...1.0.10