Local AI Bench

Release Notes(’26. 9. 12) • Version 1.1.2 New - Latest news search : with web search on, questions like "what's new / today / latest" - or about new products and terms not yet on Wikipedia - pull recent articles (title, source, date). The article titles are shown under the answer - Search API key (optional) : add a Brave Search or Tavily key in Settings to use page excerpts from real web results for more accurate answers. Mention a company's official site ("check Apple's website") and it looks there first. The key stays in this device's Keychain, and news / Wikipedia search keep working without one - Thinking mode toggle : turn reasoning models' thought process on or off. The same setting applies to chat and every benchmark, so model comparisons stay fair - Location (optional) : turn it on to answer "nearby / my area" questions; together with web search, "What's the weather now?" gets current conditions. Only an approximate location is used and nothing is stored - Settings screen : language, web search, thinking mode, location, context length, generation parameters and chat deletion in one place - One chat per model : the chat list shows the model actually used, and picking a different model opens a new chat - Korean items in the capability test : proverbs, honorifics, particles and common knowledge - Model library : sort by newest, with release month shown - New models : Qwen3.5, Kanana 2 (Kakao), Gemma 4 E4B, LFM2.5, Granite 4.1 -> tap "Update" to get them right away Improvements - Web search : fixed search not running with some reasoning models. Referenced document and article titles now appear under the answer - Web search accuracy : the product or name in your question is always kept in the search, and search terms no longer drift toward earlier answers - Fixed reasoning models (Qwen3.5, Gemma 4, etc.) leaking their thought process into the answer, or thinking without ever answering -> it now goes into the collapsible "Reasoning" box Stability - Fixed a rare crash when starting a new chat while an answer was still generating - Fixed cancelling a download letting several queued models download at once (always one at a time now) - Fixed a benchmark failing when re-run right after being cancelled - Live vision : fixed the camera staying on when the screen was closed during the permission prompt - Chat and benchmarks now share a single engine to avoid memory pressure (the send button is briefly disabled during a benchmark)