Apple Silicon AI

Apple Silicon AI — Your Macs Are the Cluster

Turn your Mac Studio, Mac Mini, MacBook Pro, or Mac Pro into a local Apple Silicon AI fleet. One endpoint routes LLM inference, image generation, speech-to-text, and embeddings across every Apple Silicon device on your network.

No cloud APIs. No GPU rentals. No Docker. Your Apple Silicon M1/M2/M3/M4 chips with unified memory are already better inference hardware than most cloud instances — you just need software that treats them as an Apple Silicon fleet.

Why Apple Silicon for AI

Apple Silicon unified memory keeps the entire model in one address space — no PCIe bottleneck, no CPU-GPU transfer overhead. A Mac Studio with M4 Ultra and 256GB runs 120B parameter models that would need multiple NVIDIA A100s. That is the Apple Silicon advantage.

Apple Silicon Fleet Setup

1. Install on every Apple Silicon Mac

pip install ollama-herd    # Apple Silicon optimized inference router

2. Start the Apple Silicon router (pick one Mac)

herd    # starts Apple Silicon router on port 11435

3. Start the Apple Silicon node agent on every Mac

herd-node    # Apple Silicon node auto-discovers the router

That's it. Apple Silicon nodes discover the router automatically on your local network. No IP addresses to configure, no config files. For explicit connection, use herd-node --router-url http://<router-ip>:11435.

How Apple Silicon routing works

MacBook Pro (M3 Max, 64GB)  ─┐
Mac Mini (M4, 32GB)          ├──→  Apple Silicon Router (:11435)  ←──  Your apps
Mac Studio (M4 Ultra, 256GB) ─┘

The Apple Silicon router scores each device on 7 signals and routes every request to the best available Mac — thermal state, memory fit, queue depth, and more.

Apple Silicon LLM Inference

Run Llama, Qwen, DeepSeek, Phi, Mistral, Gemma, and any Ollama model across your Apple Silicon fleet.

OpenAI-compatible API (Apple Silicon backend)

curl http://localhost:11435/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama3.3:70b",
    "messages": [{"role": "user", "content": "Explain Apple Silicon unified memory architecture"}]
  }'

Ollama-compatible API

curl http://localhost:11435/api/chat \
  -d '{"model": "qwen3:32b", "messages": [{"role": "user", "content": "Compare Apple Silicon M4 vs M3 for AI inference"}]}'

Apple Silicon Python Client

from openai import OpenAI
# Apple Silicon inference client
apple_silicon_client = OpenAI(base_url="http://localhost:11435/v1", api_key="unused")
apple_silicon_response = apple_silicon_client.chat.completions.create(
    model="deepseek-r1:70b",
    messages=[{"role": "user", "content": "Optimize this function for Apple Silicon"}]
)

Apple Silicon Image Generation (mflux)

Generate images using MLX-native Flux models. Runs natively on Apple Silicon — no CUDA, no cloud.

curl http://localhost:11435/api/generate-image \
  -d '{"prompt": "Apple Silicon Mac Studio rendering AI art, photorealistic", "model": "z-image-turbo", "width": 512, "height": 512}'

Apple Silicon image generation performance:

• Mac Studio M4 Ultra: ~5s at 512px, ~14s at 1024px
• MacBook Pro M3 Max: ~7s at 512px, ~18s at 1024px
• Mac Mini M4: ~12s at 512px, ~30s at 1024px

Apple Silicon Speech-to-Text (Qwen ASR)

Transcribe audio locally on Apple Silicon using Qwen3-ASR via MLX. Meetings, voice notes, podcasts — no cloud, no Whisper API costs.

curl http://localhost:11435/api/transcribe \
  -F "file=@apple_silicon_meeting.wav" \
  -F "model=qwen3-asr"

Supports WAV, MP3, M4A, FLAC. ~2s for a 30-second clip on Apple Silicon M4 Ultra.

Apple Silicon Embeddings

Embed documents across your Apple Silicon fleet using Ollama embedding models (nomic-embed-text, mxbai-embed-large, snowflake-arctic-embed).

curl http://localhost:11435/api/embed \
  -d '{"model": "nomic-embed-text", "input": "Apple Silicon unified memory architecture for AI inference"}'

Batch thousands of documents across Apple Silicon nodes instead of bottlenecking on one Mac.

Apple Silicon Fleet Monitoring

Dashboard

Open http://localhost:11435/dashboard — see every Apple Silicon Mac in your fleet: models loaded, queue depth, thermal state, memory usage, and health status.

Apple Silicon Fleet Status API

curl http://localhost:11435/fleet/status

Returns every Apple Silicon node with hardware specs, loaded models, image/STT capabilities, and health metrics.

Apple Silicon Health Checks

curl http://localhost:11435/dashboard/api/health

16 automated checks: offline Apple Silicon nodes, memory pressure, thermal throttling, VRAM fallbacks, error rates, and more.

Recommended Models by Apple Silicon Hardware

The Apple Silicon router's model recommender analyzes your fleet hardware and suggests the optimal model mix: GET /dashboard/api/model-recommendations.

Full documentation

• Agent Setup Guide — complete Apple Silicon setup for all 4 model types
• Configuration Reference — all 44+ environment variables
• API Reference — all endpoints with request/response schemas
• Troubleshooting — common Apple Silicon issues and fixes

Guardrails

• No automatic downloads: Apple Silicon model pulls are always user-initiated and require explicit confirmation. Downloads range from 2GB to 70GB+ depending on model size.
• Model deletion requires confirmation: Never remove models from Apple Silicon nodes without explicit user approval.
• All Apple Silicon requests stay local: No data leaves your local network — all inference happens on your Apple Silicon Macs.
• No API keys: No accounts, no tokens, no cloud dependencies for your Apple Silicon fleet.
• No external network access: The Apple Silicon router and nodes communicate only on your local network. No telemetry, no cloud callbacks.
• Read-only local state: The only local files created are ~/.fleet-manager/latency.db (Apple Silicon routing metrics) and ~/.fleet-manager/logs/herd.jsonl (structured logs). Never delete or modify these files without user confirmation.