MLX Models

Qwen 3.5, GPT OSS, Ministral — run open-source LLMs locally on Apple Silicon.

Local AI lets you run powerful open-source LLMs directly on your Mac for AI text processing — completely free, fully private, no internet required. Models are optimized for Apple Silicon with Metal GPU acceleration and managed directly inside VivaDicta with one-click download.

Two Runtimes, One List

Local AI covers two runtimes, and one model list spans both. Each model card is tagged with the runtime that runs it:

• MLX — the open-source catalog, in 4-bit builds optimized for Apple Silicon.
• LiteRT — Google's Gemma 4 builds (E2B, E4B and 12B).

Loading a model in one runtime unloads the other — holding two multi-gigabyte models at once buys nothing. VivaDicta recommends two models for your specific Mac based on how much memory it has, one per runtime.

When Models Load

Nothing loads at launch. The selected model warms when you start a recording or open the text-enhancement HUD, so the load hides behind you speaking or picking a preset. It then stays in memory for later requests.

macOS memory pressure frees it automatically, and the next request reloads it from disk — so the provider stays connected even while unloaded. Generation never starts a download on its own: a model that is not on disk reports that plainly instead of silently fetching several gigabytes.

Output streams as the model writes it, which matters most on the larger models.

Qwen 3.5 — Latest Models

VivaDicta includes the newest Qwen 3.5 model family from Alibaba — one of the strongest open-source LLM families available. Qwen 3.5 delivers excellent quality for text processing, translation, summarization, and coding tasks.

Other Available Models

Quantization Variants

Most models offer three quantization levels — choose based on your available RAM and quality needs:

• Base (4-bit) — smallest download, fastest inference, slightly lower quality. Best if RAM is tight.
• Med (6-bit) — balanced quality and speed. Recommended for most users.
• High (8-bit) — best quality, closest to the original model. Requires more RAM.

Getting Started

1. Open VivaDicta settings → AI Provider → Local AI.
2. Browse models organized by your Mac's RAM tier (16 GB / 32 GB / 48+ GB). A Recommended badge highlights the best model for your hardware.
3. Click Download — models are downloaded once and stored locally.
4. Select the downloaded model as your active AI provider.
5. Done — all AI processing now runs locally on your Mac.

Managing Models

• The model browser shows your Mac's memory, storage used by models, available disk space, and download count.
• Delete models you no longer need to free up disk space.
• Switch between downloaded models anytime — no re-download needed.

System Requirements

• Apple Silicon Mac (M1 or later) — required for MLX.
• RAM — varies by model (see tables above). The model browser shows which models fit your Mac.
• Disk space — models range from ~2 GB to ~20 GB depending on size and quantization.

MLX vs Cloud AI

For a comparison with Apple Intelligence and Ollama, see Local AI Processing.