Local AI Benchv1.1.4

Released October 5, 2026Major

iOS

The latest major update introduces significant technical expansion through new model support and granular performance monitoring tools. As no window analysis data is available to measure user response, the impact remains neutral. Future assessments will focus on whether these additions improve the stability of local model execution.

What changed

PerformanceAdded live performance metrics showing generation speed and elapsed time
UXAdded search progress indicators for web-enabled queries
New featureAdded open source reference chips to link to original documents
UXAdded one-tap copy button for AI answers
UXAdded reasoning preview with elapsed time and non-disruptive expansion
New featureAdded status tracking for auto benchmarks and capability tests
UXAdded manual cleanup tool for unused model files in settings
ContentAdded support for new models including MiniCPM5, Qwen3.5, and Ministral 3
PerformanceUpdated inference engine to latest MLX version
PerformanceImplemented memory check warning to prevent crashes when loading large models

Reception

50/ 100 impact scoreNeutral

What worked

Performance
The implementation of memory check warnings addresses potential stability issues when loading large models.

What didn't

New feature
no evidence
Insufficient user feedback exists to evaluate the adoption of open source reference chips or auto benchmark status tracking.

Marlvel takeaways

Monitor user engagement with the new manual cleanup tool to determine if it reduces storage-related friction for local model management.
Assess the impact of the updated inference engine on generation speed metrics to confirm performance gains for the new model support.
Evaluate the utility of the reasoning preview expansion to ensure it provides transparency without cluttering the interface.

Original release notes

Release Notes(’26. 10. 3)
• Version 1.1.4

New
- Live performance : while an answer is generating, speed (apsed time update in the performance bar
- Search progress : with web search on, you can see "preparing the search query -> searching for the actual query", so you know what the AI is
looking up
- Open sources : referenced documents and articles appear as chips under the answer - tap one to open the original
- Copy button : copy an answer with one tap, without openin
- Reasoning preview : while a reasoning model is thinking you see the elapsed time and its latest line, and expanding it no longer pushes the
screen away
- Benchmark progress : the auto benchmark and capability test show each item as pending, running or done, with its result
- Live model downloads : downloaded size, speed and time lel file-finishing step is called out separately
- Clean up unused model files : remove leftover files of models no longer in the list from Settings -> Data
- New models : MiniCPM5, Qwen3.5 0.8B, Ministral 3, LFM2.5 anbeige 4.2

Improvements
- Follow-up requests : with web search on, requests like "write the problem and its solution" or "summarize that" no longer get answered from unrelated search results -> they continue the conversation
- You can scroll up while an answer is still generating, and jump back to the latest with a button
- Models you add yourself now show their size and memory rel download size
- Updated to the latest inference engine (MLX)
- Fixed long model names overlapping the icons at the top

Stability
- Fixed the app closing when loading a model larger than the device's memory -> you now get a message before loading
- Fixed a model failing to load when it was re-downloaded adownload
- Fixed the typing indicator spinning forever after stopping a reply
- Fixed errors during a search being hidden