› cavi-ai/

/solutions/local-models · Apple Silicon

Run capable models locally with an evidence-backed path from discovery to wiring.

Find MLX models that fit your Apple Silicon host, verify them against an installed runtime, and wire a selected model into an agent workflow.

/path

How the path works

Start with the host, verify the candidate, then change configuration.

  1. 01

    Discover

    Query current MLX models and compare their real download size with the host memory.

  2. 02

    Verify

    Test a candidate against an installed local runtime before recommending it.

  3. 03

    Wire

    Emit configuration for the selected runtime target with confirmation before changes.

/start

Start with MLX Agent

macOS on Apple Silicon with Python 3.9 or newer.

pluginv0.7.0 · preview

MLX Agent

Discover, adopt, and safely wire local MLX models on Apple Silicon.

Open
skillprovided by MLX Agent

mlx-scout

Discover MLX-optimized models suited to this Apple Silicon host, routed by role.

Open

/models

Models we converted for Apple Silicon

Published on Hugging Face (opens in new tab) at the revision each row links.

Models converted for Apple Silicon
ModelTypeBuildMeasuredBase model
Audio8-ASR-Infinite-MLX-8bit (opens in new tab)hf download cavi-ai/Audio8-ASR-Infinite-MLX-8bitSpeech recognitionStreaming · English, MandarinMLX 8-bit5.29 GB7.74% WER73 LibriSpeech clips · 5.7 GB peakEdge0/Audio8-ASR-Infinite (opens in new tab)rev d3eddf9 (opens in new tab) · port source (opens in new tab)Apache-2.0 weights · MIT code
Audio8-ASR-Infinite-MLX-4bit (opens in new tab)hf download cavi-ai/Audio8-ASR-Infinite-MLX-4bitSpeech recognitionStreaming · English, MandarinMLX 4-bit3.75 GB8.35% WER73 LibriSpeech clips · 4.2 GB peakEdge0/Audio8-ASR-Infinite (opens in new tab)rev 4fe4166 (opens in new tab) · port source (opens in new tab)Apache-2.0 weights · MIT code
laya-MLX-8bit (opens in new tab)hf download cavi-ai/laya-MLX-8bitText classificationRouting, moderation · English, multilingualMLX 8-bit1.62 GB59/59 answersmatch fp32 · 12–46 ms per requestconvaiinnovations/laya (opens in new tab)rev 63ebab6 (opens in new tab) · port source (opens in new tab)Apache-2.0 weights · MIT code

Measured on an Apple M5 Max; baselines, load steps, and requirements are in each model card. CAVI is not affiliated with Edge0 or convaiinnovations.