v1.0.15

AstraLM v1.0.15 Release Notes

• Optimized compact action button layout in model cards for all screen widths.
• Perfectly aligned Resume and Cancel/Delete controls without horizontal wrapping or overflow.
• Multi-stream parallel download and background persistence engine.

v1.0.14

AstraLM v1.0.14 Release Notes

• Fix .part cleanup bug on app restart to preserve partial downloads.
• Add interactive Resume button on paused and interrupted downloads.
• Fix RandomAccessFile seek write mode for multi-stream chunk slices.
• Fix CancelToken reuse on resumed downloads to eliminate stuck 0% states.
• Fully unkillable background download engine with Android ForegroundService dataSync.

v1.0.13

AstraLM v1.0.13 Release Notes

• Parallel 4-stream resumable model downloader with sparse chunk allocation.
• Android ForegroundService dataSync support for unkillable background downloads.
• Crash and power-off state persistence via .part.meta chunk recovery.
• Zero-flicker model browser with GPU layer caching and isolated card rebuilds.
• Automatic binary GGUF header validation on completed models.
• Clean Phosphor Bold interface icon system.

AstraLM v1.0.12

Download Engine Overhaul

• Replaced legacy Android DownloadManager with a direct high-performance in-app socket streaming engine, eliminating the system-level 1 Mbps bandwidth ceiling on Xiaomi HyperOS and Samsung devices.
• Added dynamic foreground WifiLock and Partial WakeLock integration to keep connections running at full wire speed with zero throttling.
• Added smooth progress debouncing to eliminate UI thread latency during high-speed multi-megabyte transfers.
• Added byte-range resumption support for interrupted downloads.

AstraLM v1.0.11

Performance & Download Engine

• Added high-performance WifiLock and Partial WakeLock to prevent network throttling and socket termination on Xiaomi HyperOS, Redmi, and Samsung OneUI devices during large model transfers.
• Added battery optimization exemption handling to allow uninterrupted background model downloads.
• Integrated direct edge CDN resolution (?download=true) for all large GGUF weights.

Dynamic Model Catalog & OTA Sync

• Implemented On-The-Air (OTA) model catalog synchronization: model lists, quantizations, and download endpoints now update dynamically in the background without requiring APK reinstalls.
• Added Kimi Moonlight 16B-A3B Fast (IQ2_XXS - 5.50 GB) optimized quantization for faster downloads and lower memory footprint on 6GB/8GB RAM devices.
• Updated Flagship Kimi Moonlight 16B-A3B (Q3_K_S - 7.65 GB) telemetry and CDN mirrors.

Interface & Design

• Integrated Phosphor Bold icon system across all navigation, input, and model management views.
• Upgraded Fullscreen Canvas with unified markdown rendering and live HTML preview.
• Updated sovereign brutalist monochrome website design with dynamic GitHub release resolution.

AstraLM v1.0.10

What's New in AstraLM v1.0.10

• Android 16 & Motorola G56 Compatibility: Fixed Scoped Storage / MediaStore permission exceptions in DownloadManager by routing downloads to permission-exempt app-specific directories.
• Download Recovery on Reopen: Fixed stalling and frozen progress bars when reopening AstraLM while downloads are active.
• Smoothed Speed & Accurate ETA: Added Exponential Moving Average (EMA) smoothing for download speeds and fixed negative/NaN ETA durations.
• Silent Background OTA Updates: AstraLM automatically checks and downloads release updates in the background, ready for instant 1-tap installation.
• Tecno & MediaTek Flash Glitch Fix: Fixed screen flashing/flicker when navigating to the Models view by setting non-opaque route composition and solid background layers.

AstraLM v1.0.9

AstraLM v1.0.9 Release Notes

Release: v1.0.9 (Build 9)
Package: com.prasun01.astralm
Target: Android 8.0+ (ARM64-v8a)

---

Important Notice: Clean Package ID Migration

In this release cycle, AstraLM has officially migrated to its permanent, independent package identifier (com.prasun01.astralm).

• Upgrading from v1.0.7 or earlier: If you have an older version installed, Android will install v1.0.9 as a new standalone application rather than overwriting the previous app. Please uninstall the old version from your device.
• Future Updates: This is a one-time migration. All future updates (v1.1.0, etc.) from v1.0.9 onward will update seamlessly in-place without duplicate apps or manual uninstalls.

---

Highlights in v1.0.9

• Model Download Engine Overhaul: Implemented native Android DownloadManager bridges and a resilient, auto-resuming streaming engine. If a network connection drops midway, downloads automatically resume from the last byte without restarting.
• 100% Verified Model Catalog: Replaced all gated model endpoints with direct, verified public HTTP 200 URLs across the entire 37-model catalog (Qwen 2.5, DeepSeek-R1, Gemma 2, Llama 3.2, Stable Diffusion).
• Calibrated RAM Safety Evaluator: Fixed transformer KV cache memory formulas in DeviceInfoService to prevent false high-RAM warning dialogs on 1B to 3B parameter models.
• Side Drawer Performance Optimization: Replaced frosted glass blur filters with a crisp, solid surface container for 120 FPS drawer slide transitions.
• One-Handed Ergonomics: Moved primary navigation actions to the bottom thumb zone and integrated a 1-tap New Chat action in the top AppBar.
• Gemini Multimodal Fix: Resolved payload parsing issues affecting Google Gemini image models.

---

Assets

• Binary: AstraLM-v1.0.9.apk (76.5 MB)
• Package Name: com.prasun01.astralm
• Architecture: arm64-v8a

AstraLM v1.0.7 — Redesigned Walkthrough & Smart Online Models

[1.0.7] - 2026-08-25

Added

• Redesigned Welcome Walkthrough: Implemented an obsidian-themed 3-step onboarding flow with borderless elements, fluid page transitions, and structured initial paths:
  • Step 1: On-device private inference overview.
  • Step 2: Universal cloud ecosystem and client-side encryption.
  • Step 3: Getting started actions (Download starter local model or connect cloud API key).
• Smart Online Model Filtering: Intelligent filtering across cloud provider endpoints (Google Gemini, OpenAI, DeepSeek, OpenRouter, NVIDIA NIM):
  • Filters out embedding models, speech/audio endpoints (TTS, Whisper), moderation endpoints, and legacy completion models.
  • Prioritizes top chat, reasoning, and multimodal models (DeepSeek-R1, Gemini 2.5 Pro/Flash, Claude 3.5 Sonnet, GPT-4o).
• Online Models FAQ & Setup Documentation: Added an expandable accordion guide inside the Online Models view explaining API keys, recommended providers, privacy guarantees, and custom server configuration.
• Monochrome Provider Logos: Added custom monochrome vector assets for OpenRouter, OpenAI, DeepSeek, Google Gemini, Anthropic, NVIDIA NIM, Groq, and Stability AI.
• Android-Optimized Local Models: Added verified low-memory local models to the catalog (Llama-3.2-1B LiteRT-LM, SmolLM2-360M, SmolLM2-1.7B, Ministral-3B, Qwen2.5-0.5B, DeepSeek-R1-Distill-Qwen-1.5B).
• Background Download Notifications: Configured Android DownloadManager with completed visibility and a dedicated notification channel (astralm_model_downloads).

Changed

• Local Model Filter Bar: Migrated to a single-row horizontal scrolling chip layout with smooth left and right fading edge masks.
• Filter Registry Update: Fixed category routing and item count calculations for LiteRT, GGUF, and Reasoning models in ModelController.
• UI Element Cleanup: Removed grey outlines and unnecessary icon clutter across buttons and filter pills.

---

AstraLM v1.0.6

🚀 AstraLM Release v1.0.6 — Changelog & Security Audit

Release Tag: v1.0.6
Build: 1.0.6+6
Target Platform: Android (arm64-v8a)
Binary Location: build/app/outputs/flutter-apk/app-release.apk (76.4 MB)
Package Name: com.orailnoor.privatelm

---

🔒 Security Audit & Penetration Review Summary

Security Domain	Status	Details
Android Manifest Hardening	✅ PASS	All background services explicitly marked android:exported="false". Watchdog/Boot receivers cleanly stripped to prevent unauthorized component invocation.
Cleartext Traffic & Transport Security	✅ PASS	Strict HTTPS across all cloud endpoints (Anthropic, OpenAI, Google Gemini, OpenRouter, DeepSeek, Stability AI). Zero unencrypted HTTP endpoints.
API Key Storage & Memory Hygiene	✅ PASS	Cloud API keys stored locally in private application sandbox (Hive Encrypted/Local Box). No hardcoded credentials or key leakage in logs/crash reports.
WebView Sandboxing	✅ PASS	Canvas Web preview scoped with isolated DOM context, sanitized script injection, and eager gesture separation.
Native Certificate & Package Signing	✅ PASS	Signed with official release RSA-2048 keystore (astralm-release.jks) using APK Signature Scheme v3.
Static Code Analysis	✅ PASS	All critical linter errors, unhandled async context warnings, and dead imports resolved across the codebase.

---

🌟 What's New in v1.0.6

🎨 Visual & UX Refinements

• Claude & ChatGPT-Style Instant Launch:
  • Zero artificial splash delay screens.
  • Native launch window displays centered brand icon and transitions immediately into the chat workspace.
• Dynamic & Witty Thinking Indicators:
  • Progressive status indicator during model reasoning (Thinking… ➔ Cooking… ➔ Let him cook… ➔ Thinking harder… ➔ bro, your cpu is slow 💀).
  • Rock-steady, jitter-free cross-fade animations replacing layout-shifting slide transforms.
• Full-Screen Borderless Canvas:
  • Removed top header bars, document titles, and bottom background bars.
  • Floating borderless action buttons with subtle elevation shadows.
  • Smooth top and bottom vertical gradient fading masks.
• Zero Touch-Latency Web Preview:
  • Hardware-accelerated EagerGestureRecognizer integration for 0ms touch and scroll latency on live HTML/JS previews.

🌐 Cloud & Image Generation

• Stability AI v2beta API Fix:
  • Auto-upgraded model resolution to sd3.5-large-turbo, resolving 400 Bad Request: invalid value for 'model' errors.
  • Integrated endpoint routing for sd3, core, and ultra.
• OpenAI DALL-E 3 Support:
  • Direct routing for dall-e-3 / dall-e-2 image generation with native base64 in-chat preview rendering.
• Expanded Token & Context Windows:
  • Bumped default max tokens for online cloud models to 8,192 tokens.
  • Expanded document attachment capacity to 500,000 characters (~125k tokens).

☕ Developer Support

• Discreet Sidebar Support Tile:
  • Integrated a clean, non-intrusive Buy Me a Coffee support button in the drawer footer linking directly to the creator's page.

---

📦 APK Signing & Certificate Details

Signer #1 Certificate DN: CN=AstraLM, OU=Mobile, O=AstraLM, L=SanFrancisco, ST=CA, C=US
Signature Scheme: APK Signature Scheme v3
SHA-256 Digest:   aac94d12645b5ec7b5234d82c7755bab5a41236eb3a6ec0f76dc3ed37385a0aa
SHA-1 Digest:     f683d0e20c6c95623450084712b8cec470f34ce6
Key Algorithm:    RSA 2048-bit

AstraLM v2.0.0

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.