Built-in browser and local inference updates

Built-in browser

Browse websites inside Luxi. You and the AI share the same pages, with multiple tabs and a Pause AI control that lets you take over. Website sign-ins are isolated by conversation; passwords and file uploads remain manual. Website file downloads are not yet supported, and tabs are not restored across app restarts.

Conversations and context

The conversation runtime has been upgraded, with support for updated session and configuration formats and refreshed tool and sidebar integration. The context usage indicator now appears beside the model selector.

Local inference engine

Long-input prefill now uses memory-bounded chunks and automatically calibrates eligible devices. This update adds a hardware-gated attention path for supported M4 GPUs and fixes part of the multi-step verification computation on non-NAX GPUs. The first eligible model load may take longer while calibration runs; subsequent loads reuse its results. These changes do not establish that every device-specific incorrect-answer issue is resolved.

Known issue

During local acceptance, Luxi Pro once deviated from an explicit instruction and invoked unrelated tools. Subsequent tests did not reproduce it, but the cause remains unconfirmed; this engine update should not be interpreted as a fix for that behavior.

Upgrading

Existing Luxi Pro/Lite MLX models do not need to be downloaded again. The 8K default reasoning budget, UTF-8 guard, resumable model downloads and Pro MTP remain available. GGUF users still need the corresponding MLX models.

This version includes a session-format upgrade. Quit Luxi and back up its data before upgrading. To downgrade, restore the complete pre-upgrade backup; older clients may not be able to read upgraded conversations.

This release is for Apple Silicon Macs running macOS 14 or later.