Local inference

Task-specific commands run compatible local model directories through declared backends. Choose a model with Intake, install its backend if needed, and finish conversion before running these commands.

These commands execute directly; they do not use the preview/confirmation contract of downloads and lifecycle starts. Review inputs, output paths, and resource needs.

Speech

python3 scripts/mlx-agent convert transcribe --path /absolute/local/asr-model --audio ./recording.wav --json
python3 scripts/mlx-agent convert speak --path /absolute/local/tts-model --text "Hello." --out /absolute/output/speech.wav --json

Transcription accepts an audio file and optional --language. Speech output requires a new absolute .wav path; voice, speed, and language depend on the model.

Music

python3 scripts/mlx-agent convert music --path /absolute/local/music-model --caption "A gentle piano melody" --lyrics '[instrumental]' --out /absolute/output/music.wav --duration 15 --steps 30 --seed 42 --json

The isolated audio backend supports MiniMax Music 3 from local weights. Duration is a maximum request greater than zero and up to 360 seconds; steps range from 1 to 30. Lyrics may include section tags. Output is a new mono or stereo PCM WAV. Reported duration, generation/load time, real-time factor, and memory do not measure listening quality.

The checkpoint needs a real local tokenizer at tokenizer/tokenizer.json or root tokenizer.json; a synthetic tokenizer fallback is refused. Already-exported quantized checkpoints can be converted from a local --source-path, using the model's layer policy while preserving tokenizer and scheduler assets.

Vision and classification

python3 scripts/mlx-agent convert describe --path /absolute/local/vision-model --image ./image.png --prompt "Describe this image." --json
python3 scripts/mlx-agent convert describe --path /absolute/local/vision-model --video ./clip.mp4 --prompt "Summarize this clip." --max-tokens 256 --json
python3 scripts/mlx-agent convert decide --path /absolute/local/classification-model --request ./request.json --json

Describe supports image or video input, bounded answer length, and media sampling controls. Decide reads a JSON request with state and questions fields for a compatible classification model. A text-generation model directory is not interchangeable with these task-specific models.

Image and video generation

python3 scripts/mlx-agent convert generate --path /absolute/local/image-model --prompt "A green apple on a table" --out /absolute/output/image.png --json
python3 scripts/mlx-agent convert video --path /absolute/local/video-model --prompt "A green apple rolling across a table" --out /absolute/output/video.mp4 --json

Use a new absolute output path ending in .png or .mp4. Width, height, steps, frame counts, and task support depend on the backend/model. Command help gives bounds and defaults:

python3 scripts/mlx-agent convert video --help
python3 scripts/mlx-agent convert describe --help

Use --timeout to bound execution. Keep sensitive inputs and generated media private. Synthetic adoption probes do not describe the privacy properties of user-supplied inference inputs or their logs.