Inferencer v1.2 with token inspection

Version 1.2 is now available for download on the Mac App Store. This update introduces major memory savings and improvements to the token inspector.

Token inspector

We’ve made improvements to the token inspector interface, not just visual, as you can now directly select alternative tokens and have the model automatically regenerate the response.

Memory savings

Depending on your usage you may find a lot more free memory available in version 1.2. In the screenshot below you can see both v1.1 and v1.2 running side by side, with v1.2 bringing over a GB in savings allowing for larger models and larger context windows.

Inferencer is written in native code, which allows us to optimise our codebase to use less resources than typical Chromium-based web-app wrappers. It also runs completely sandboxed for safer inferencing with models that utilise custom python tokenisers.

More improvements

• Choose the number of token probabilities to generate
• Manually load and unload the model as automatic timeout settings
• Prompt processing progress indicator
• Markdown images can now be copied
• Dark/light appearance selection
• Support /think tag in markdown
• Select and delete multiple conversations (hold shift key)

Preview features

We’ve added a section in the settings page to allow for upcoming features.

• Token entropy - allowing you to instantly see low confidence tokens.
• Token exclusion - allowing you to exclude certain tokens from generations (e.g. foreign language characters).

Public roadmap

You can now submit and vote features to be added to Inferencer in our public roadmap.

Demo Videos

We have two demonstration videos to share.

• The first is a preview edition of Inferencer running on the new iPhone 17 Pro: https://youtu.be/mrBkWuXZTdA
• The second is DeepSeek V3.1 Terminus running on the latest version of Inferencer: https://youtu.be/kts098EL2PQ

If you have any suggestions you’d like to share in private, I’d love to also hear them.

Thanks again for your support, and as a reminder, all AI processing is done offline, directly on your device. No telemetry, no background "update" checks.