Skip to content

AstraLM v2.0.0

Choose a tag to compare

View all tags
@Prasun01 Prasun01 released this 24 Aug 14:35
· 78 commits to main since this release
v2.0.0
086965a

AstraLM — Release Notes & Changelog

Important

> ### ⚠️ Crucial Upgrade Note for Existing Users:
> Due to a production signing key upgrade and security overhaul in v2.0.0, you must uninstall the previous version from
your device before installing this APK.

🚀 Version Overview

This release delivers a comprehensive architectural overhaul, strict monochrome styling, a complete icon system migration to Phosphor Bold, zero-lag Canvas Workspace, live unbuffered online model streaming, robust multi-model memory management, and Samsung Galaxy Mali/Exynos GPU hardware stabilization.

✨ Key Features & Improvements

1. 🎨 Design & UI Modernization

  • Strict Monochrome Palette: Standardized pure black (#000000), neutral dark greys (#101010, #141416, #181818), and crisp white typography across all screens.
  • Borderless Modern Aesthetics: Removed all grey borders, harsh card outlines, and legacy dividers across Settings, Popovers, and Dialogs.
  • Phosphor Bold Icon System: Replaced all Material icons with genuine Phosphor Bold icons (PhosphorIconsBold.*).
  • Smooth Gradient Fade Transitions:
    • Expanded conversation drawer session list fade seamlessly above the Settings footer button.
    • Enhanced bottom chat gradient fade (64px) behind the floating action bar.

2. ⚡ Live Unbuffered Online Model Streaming

  • Zero-Latency Real-Time SSE Streaming: Replaced buffered HTTP calls with unbuffered dart:io HttpClient socket streams (Cache-Control: no-cache).
  • Google Gemini Streaming: Integrated native streamGenerateContent?alt=sse for real-time word-by-word streaming.
  • Anthropic Claude & DeepSeek-R1 Support: Added native SSE delta streaming and real-time reasoning token (reasoning_content) visibility.
  • Smooth Word-by-Word Fallback: Any batch endpoint smoothly renders with adaptive progressive word streaming and soft AnimatedOpacity fade-in.

3. 🛡️ Single-Session Incognito Mode

  • Zero-Persistence Privacy: Messages in Incognito mode are never saved to Hive or disk.
  • Auto-Deletion Lifecycle: Incognito is strictly confined to the active chat session. Toggling off, closing the chat, tapping New Chat (+), or switching conversations automatically destroys all private messages from memory.

4. 🧠 Memory & Inference Engine Optimization

  • Sequential Local Model Switching: Selecting a new local model or cloud provider proactively triggers a full unloadModel() and GPU/VRAM memory flush before instantiating the new engine.
  • Samsung Galaxy / Mali GPU Protection:
    • Added dynamic context size clamping (1536–2048 tokens) for Gemma / LiteRT models on Mali-G68 / Exynos SoCs.
    • Implemented automatic CPU Safe Fallback in InferenceAndroidService if GPU OpenCL compilation runs out of memory on Mali.
  • Vision Model Fixes:
    • Fixed Qwen2-VL-2B catalog validation threshold preventing false "Incomplete Model File" errors.

5. 📝 Canvas Workspace 2.0

  • Instantaneous Tab Switching: Replaced TabBarView with IndexedStack to keep InAppWebView live in memory with 0ms lag.
  • Accurate Code Extraction: When opening "Open in Canvas", extraneous AI conversational preamble is cleanly stripped, extracting only the clean markdown/code.
  • Clean Borderless Floating Cards: Polished document styling, syntax highlighting, and action bars.

🛠️ Summary of Changed Files

File Changes
lib/views/chat_view.dart Phosphor icons, Settings drawer footer, expanded gradient fades, quick model switcher
lib/views/settings_view.dart Nested sub-settings, strict monochrome design language, zero grey borders
lib/services/cloud_service.dart Unbuffered HttpClient SSE streaming for Gemini, Claude, OpenAI, DeepSeek, Kimi
lib/services/device_info_service.dart Mali/Exynos calibration and safe context limits for Gemma models
lib/services/inference_android.dart LiteRT GPU failure recovery & CPU safe mode fallback
lib/controllers/chat_controller.dart Single-session incognito lifecycle, stop-generation state, progressive stream animation
lib/controllers/model_controller.dart Eager model unloading, robust GGUF validation, solid 100% opaque notifications
lib/widgets/canvas_workspace.dart 0ms lag IndexedStack tab switcher, Phosphor icons, document view
lib/core/icons.dart Centralized Phosphor Bold icon registry
lib/core/constants.dart Updated Qwen2-VL-2B metadata & cloud endpoints

Important

> ### ⚠️ Crucial Upgrade Note for Existing Users:
> Due to a production signing key upgrade and security overhaul in v2.0.0, you must uninstall the previous version from
your device before installing this APK.

Assets 3
Loading