Zum Inhalt springen
TakoAPI
Zurück zu den Skills

mlx-local-inference

von bendusy

DevOps & Cloud
329Downloadzahlen laut clawskills.sh
1Sterne
0Aufrufe
0 LikesMelde dich an, um diesen Skill zu liken
—Noch nicht bewertet
Deine Bewertung

Kurzbeschreibung

Full local AI inference stack on Apple Silicon Macs via MLX.

Beschreibung

Was dieser Skill macht

Runs a full AI inference stack locally on Apple Silicon Macs using the MLX framework. Covers LLM chat (Qwen3-14B, Gemma3-12B), speech-to-text via Whisper and Qwen3-ASR, text embeddings, OCR with PaddleOCR, and text-to-speech. All services expose OpenAI-compatible APIs on localhost.

Eliminates cloud API costs and keeps all data local while exposing the same OpenAI-compatible API format that most AI tooling already supports.

Wann du ihn einsetzt

  • Transcribing Cantonese audio without uploading to cloud services
  • Running privacy-sensitive LLM queries entirely on-device
  • Extracting text from scanned documents or images via OCR
  • Building local semantic search with on-device text embeddings
  • Auto-transcribing a folder of meeting recordings in the background

Beispiel-Workflow

EINGABE

User asks: Transcribe and clean up the Cantonese meeting recording at ~/Downloads/meeting.m4a

AGENT
  1. 1

    Copy meeting.m4a into ~/transcribe/ to trigger the daemon

  2. 2

    Daemon detects the file within 15 seconds and sends it to Qwen3-ASR on port 8788

  3. 3

    ASR output is saved as meeting_raw.md with the raw transcript

  4. 4

    Qwen3-14B reads the raw transcript, corrects homophones, adds punctuation, and preserves Cantonese characters

  5. 5

    Corrected file is moved to ~/transcribe/done/meeting_corrected.md

AUSGABE

meeting_corrected.md in ~/transcribe/done/ with a clean, punctuated Cantonese transcript

Voraussetzungen

  • Apple Silicon Mac (M1, M2, M3, or M4)
  • Python packages: mlx, mlx-lm, mlx-audio, mlx-vlm

Installation

clawhub install bendusy/mlx-local-inference

Führe den CLI-Befehl aus oder füge den Prompt in Claude Code, Cursor, Codex oder deinen Coding-Agenten ein.

Links