MLX local models and clearer startup
Updated Pro/Lite local models, startup warmup, a default 12K-token reasoning budget and simpler settings. Download MLX models when upgrading; existing data is preserved.
MLX local models
Local inference now runs on MLX. Pro uses Qwen3.6-35B-A3B-OptiQ with MTP enabled automatically; Lite uses Qwen3.5-9B in 6-bit format. Both models support text and image input.
Startup and responses
The workspace opens after model loading and initial warmup complete, with loading status shown during startup. Local models use a default 12K-token reasoning budget and consistent sampling settings.
Simpler model settings
Model cards retain the Luxi Pro/Lite names, size, status and controls while hiding the underlying model and quantization names.
Upgrading from 0.4.36
Existing GGUF models cannot run on the new engine. Follow the setup flow to download the corresponding MLX models. Existing GGUF files, conversations, settings and workspace files are preserved; old models are not automatically deleted. Downloads support resuming and integrity verification.
This release provides an Apple Silicon macOS installer and requires macOS 14 or later. Lite requires at least 16 GiB unified memory, with 24 GiB recommended; Pro requires at least 32 GiB, with 48 GiB recommended. Longer conversations remain subject to available memory.