Faster, sharper answers • Small documents are now read in full, so answers are quicker and more accurate. • Every document answer shows its source — tap to open the page it came from. You choose how it thinks • Thinking: Off, Short or Full. Each model starts with the setting that suits it best. New model screen • One card per model, picks recommended for your device, new Qwen 3.5 models. Truly offline • In airplane mode, web search is no longer offered, so the AI answers from what's on your device. Fixes • Cleaner lists and formatting, no more "message too long", better Gemma 4 tool use, many stability improvements. • Knowledge Packs and Wikipedia import have been retired.

This is the biggest update to Bastion Chat yet. VISION & IMAGE UNDERSTANDING Ask questions about photos and images. Multimodal models can now see, describe, and analyze images directly in your conversations. Image history is preserved across sessions. AGENTIC MODE Bastion Chat can now act, not just chat. Enable Agent Mode for multi-step reasoning with autonomous tool use — your AI plans, executes, and delivers results. WEB SEARCH Optionally connect to the web for real-time information. Ground answers in live results without leaving the app or compromising your privacy. WIKIPEDIA IMPORT Pull any Wikipedia article directly into your knowledge base and query it instantly with full RAG. BUILT-IN KNOWLEDGE PACKS Curated, system-managed knowledge packs (first one available: Pet Care) come pre-loaded and searchable — no setup required. SMARTER MODEL BROWSER Redesigned with collapsible sections, family grouping, inline quantization picker, and quality tiers so you always know which model fits your device best. FAVORITES Star your most-used models for instant access at the top of the model list. PER-MODEL SAMPLING PROFILES Fine-tune temperature, context length, and inference settings independently for each model. SMARTER CUSTOM MODEL IMPORT Auto-detects capabilities (reasoning, tool calls, vision) on import. No manual configuration needed. TURBOQUANT KV CACHE New TurboQuant memory modes (3.5-bit and 2-bit) let you run larger models with dramatically less RAM — without sacrificing response quality. FLASH ATTENTION Flash Attention is now enabled by default on all Apple Silicon devices, significantly reducing memory usage during long conversations and improving inference speed. FULL METAL GPU OFFLOAD All model layers now offload to the Metal GPU by default, with device-aware thread and micro-batch tuning for the fastest possible inference on your iPhone, iPad, or Mac. GEMMA 4 THINKING SUPPORT Native support for Gemma 4's thinking block format, including proper stop sequence handling and streamed thinking output. Bug fixes: MCP crash fix, Metal SET_ROWS crash on hybrid GDN models (Qwen 3.6 27B), atomic interrupt flag, download integrity checks, and Qwen context length handling.

## Major New Features: - **Agentic Search with Thinking Models** - New intelligent search using Qwen3 thinking models - Multi-step reasoning with internal thought processes - Document analysis with step-by-step logical deduction - **Quick Search with Traditional RAG** - Fast traditional retrieval-augmented generation - Works with all model types for instant results - Alternative to agentic search for simpler queries - **Search Across Documents** - NEW FEATURE: Search within or across selected documents - Flexible document filtering and scope control - Intelligent result summarization with citations - **Tool Calling System** - Built-in tools for document search, calculations, and analysis - AI can perform actions and execute tasks - Multi-step task execution with reasoning - **Word Cloud Generation** - Visual document analysis with term frequency analysis - Natural language processing for key concept identification - Document comparison and trend analysis - **MCP Integration (Model Context Protocol)** - External tool connections for web APIs and services - Tool orchestration with intelligent selection - Privacy-first design maintaining local processing - **Agentic Workflow** - Multi-step reasoning with tool execution - Block-based message architecture for complex tasks - Intelligent task planning and execution ## UI & Experience Improvements: - Enhanced feature section layouts with better spacing - Improved capability items with grid layout - Better visual hierarchy and card designs - Enhanced mobile responsiveness - Real screenshots showing actual functionality ## Technical Accuracy Updates: - Fixed model table to show only Qwen3 models have full tool support - Updated other models to show 'Limited' tool support - Corrected MCP integration image placement - Better image placement and captions ## Platform Evolution: BastionChat has evolved from a basic chat app to a complete AI agent platform with: - Revolutionary tool calling capabilities - Visual document analysis - Agentic workflows - MCP integration for external connections - Complete privacy-first design