## 2.9 - Fixed Safari Extension not being fully restored from v2.8 - Safari Extension now automatically falls back to BYOK model when encountering errors with Auto Apple Intelligence

## 2.8 ### Fixes - Fixed recommended model files failing to load correctly - Fixed progress reporting for model downloads - Fixed Safari Extension styles being incorrectly cleared and long-form pipeline issues

## 2.7 What's New - You can now customize downloading of local models. Bug Fixes - Improved selection in the detail page; text copying now works properly. Removed - Removed MLC-LLM solution. - Removed Qwen 3 0.6b from Safari Extension.

New Features - The share function now supports text input. - The share function can now automatically open the App after sharing is complete (can be disabled in settings). - Added a system translation feature that automatically translates prompts, improving language guidance. Fixes - The share function now correctly prioritizes URLs. - Auto strategy update: Switches to BYOK for content over 1792 tokens. - Local: Chunk size is now fixed at 1792. Local: Qwen3 0.6b q4 - Now automatically cleans up unnecessary "Think" tags. - Optimized mlc-llm config: Qwen3 0.6b now performs better on iPhone 15 Pro / iPad Pro m4. - Input content is now sanitized to prevent misleading the model. Removed - Removed the option to select Tokenizer calculation methods.

## 2.5 ### Fixed - Corrected the issue with Safari Extension due to template translation problems, leading to a loss of ability to obtain text content. - Corrected the issue with the App Intent (App Intent) due to transferring URL into Text, causing the app to treat URL as a text-to-be processed. - Improved internal prompts for higher model adaptability. - Corrected the issue with temperature:0, affecting only the communication quality of LLMs.

## 2.4 ### Added - Added Auto strategy mode - Provided a “Send to eisonAI” shortcut entry, now allowing URLs or text to be sent to the app ### Fixed - Moved Qwen3 0.6b to the “Lab” section; it now requires manual activation to use. - Optimized the BYOK input experience. - Fixed an issue where passing `temperature: 0` in BYOK caused some models to reject the request, resulting in a URL‑0 error. - Long‑text splitting mode works only for local models (Apple Intelligence or Qwen3 0.6b). - Translated prompts into English to reduce conflicts with system prompts that could affect output language. This addresses weaker instruction‑following capabilities of some models. - Added a context menu implemented via right‑click to support devices that do not support gesture navigation. - Restored the Safari Extension to the Qwen3 0.6b cs1k version to avoid failures on iPhone when web‑llm cannot run (cs2k version has been tested and is too difficult to start, so it defaults to cs1k). ### Removed - Removed the BYOK long‑text strategy.

## 2.1 - We fixed an issue that made the app package much larger by accidentally including the AI model twice. The model is now bundled only with the Safari extension (as intended), so the main app is smaller while still working normally. ## 2.0 What’s New in eisonAI 2.0 eisonAI 2.0 is a complete rewrite. This version rethinks how reading tools should work — not by adding more automation, but by making structure visible while keeping judgment in your hands. Major changes - New reading model: Cognitive Index™ Structure is now the primary unit of reading. Instead of summaries or rewrites, eisonAI reveals where meaning forms inside real texts — decisions, assumptions, and turning points. - Built for long-form content Stable handling of long documents (up to 15,000 tokens). No manual segmentation. No demo-only workflows. - Safari Extension integration See structure directly on the page. No copy-paste. No context switching. - Local-first architecture On-device processing by default. Sensitive reading stays private. - Reading Library with tags Save, organize, and revisit ideas over time. Your library reflects how you think, not just what you saved. - CloudKit sync Seamless synchronization across devices via iCloud. - Redesigned onboarding A new onboarding flow that explains how to read with structure, not just what buttons do. - Performance & stability Faster rendering, smoother interactions, and a simpler internal architecture.

What’s New in eisonAI 2.0 eisonAI 2.0 is a complete rewrite. This version rethinks how reading tools should work — not by adding more automation, but by making structure visible while keeping judgment in your hands. Major changes • New reading model: Cognitive Index™ Structure is now the primary unit of reading. Instead of summaries or rewrites, eisonAI reveals where meaning forms inside real texts — decisions, assumptions, and turning points. • Built for long-form content Stable handling of long documents (up to 15,000 tokens). No manual segmentation. No demo-only workflows. • Safari Extension integration See structure directly on the page. No copy-paste. No context switching. • Local-first architecture On-device processing by default. Sensitive reading stays private. • Reading Library with tags Save, organize, and revisit ideas over time. Your library reflects how you think, not just what you saved. • CloudKit sync Seamless synchronization across devices via iCloud. • Redesigned onboarding A new onboarding flow that explains how to read with structure, not just what buttons do. • Performance & stability Faster rendering, smoother interactions, and a simpler internal architecture.